Customers

How Malachyte solves retail’s cold-start problem with managed real-time AI

Mon, 10 Aug 2026 16:00:00 +0000

What’s the best way to recommend products to little-known users?

We’ve spent our careers trying to solve this problem for major companies like Spotify and Priceline, and it’s why Sidd founded Malachyte, an AI-powered ecommerce recommendation platform. These days, consumers have come to expect content that feels personalized and relevant, and online services competing for their attention have no choice but to do this exceptionally well.

Malachyte was inspired by some unique insights into how advanced AI models, and large language models in particular, could be applied in new ways to old challenges like personalization and recommendations.

As Malachyte set out to win potential customers’ business, we needed secure, scalable, reliable and, above all, leading-edge AI infrastructure to continue building the personalization algorithm we had always envisioned. By utilizing Google Cloud tools like Bigtable and Managed Service for Apache Kafka, Malachyte has been able to help some of its retailers double and sometimes even triple their sales.

This is the story of how we built it, and the ways any founder can use services like these to start deploying AI foundation models in new ways.

How Malachyte lifted sales for their users

For Malachyte, the aha moment was discovering that it could use neural networks with attention mechanisms — the same concept powering large language models — to personalize retail search and product pages. This approach is what enables LLMs to derive meaning from the relative order of items in a sequence, in their case the order of words and syllables in a sentence. When it comes to a retail website or app, what Malachyte wanted to capture was the sequence of customer interactions with the site.

What if we predicted the next thing a user wants on an ecommerce website just like LLMs predict the next word in a sentence?

A pre-GPT language model might have tried to look at a specific sequence of words or even fragments of words (what we now know of as tokens), but those earlier models wouldn’t examine what happens if the words were in the comparable order but weren’t contiguous or were re-arranged. The breakthrough came — in part through Google’s work on transformers — when LLMs gained the ability to understand complex and long-range dependencies within a sequence of items.

This more sophisticated method has delivered dramatic results — both for the proliferation of gen AI in general, and for Malachyte’s application of the technology.

To make this work in practice, Malachyte creates a vector of everything known about a visitor when they arrive on a site. Most users are visiting for the first time, so little is known about them. This is what’s known as the “cold start” problem. The trick is to use every interaction with a user to refine this vector. Each new addition to the vector, like a click or a query, does two things: it drives a prediction about the next thing the user wants, and it provides more information about the user.

Malachyte’s platform then updates the user vector and the prediction at the same time. This not only enhances the understanding of the individual user and their preferences, it also improves the overall model with the anonymized user data. With every inference, the context of both the average and the specific shopper grows.

The company further innovates by not just using attention-based neural networks but combining that with updating user profiles 100 milliseconds at time.

Malachyte’s recommendation and search agents populate the next page’s search results or recommendation carousels based on what users clicked on previous pages.

To be sure, this idea isn’t in itself new. Retailers have long used collaborative filtering recommendations systems to identify similar users and items that required massive sets of interaction history. These models typically required a lot of data, including third-party cookie-based profiles and demographics.

By focusing on the sequence of interactions in a session, retailers can achieve far more personalization — with less required data or spend — than by focusing only on a user’s profile. As a bonus, retailers can now offer their users more privacy by not relying on long-term cookie data.

This works because of the model structure and multimodal vectors that encode everything they know about a user, including browser data, click history and searches. The output, too, is multimodal: The same model can be applied to on-site search product pages, category pages, and add-to-cart carousels.

To make this work, each product in the catalog is embedded into the same space as the user vector, which gets updated and subsequently moves the vector closer to relevant products and further from those that aren’t. The neural network computing the embedding is being continuously trained across retailers who work with Malachyte, improving the quality for everyone. The system effectively becomes a data cooperative with each retailer's user helping make the model smarter for everyone.

A user session represented as a vector in a space of products.

To make this delivery for every user at every inference in 100 milliseconds, Malachite found real benefits in building onGoogle Cloud’s real-time AI stack.

With this system, every behavioral event streams into a Managed Service for Apache Kafka cluster. Rather than queuing for a future training job, each event immediately becomes an update to the user’s profile in Bigtable. The Kafka cluster allows the customer’s front-end to persist, so the user session signals quickly with little worry about how they fit into the user vector.

Bigtable allows Malachyte’s services to look up and update the right user vectors, and it and Kafka operate at the order of 10 milliseconds per step, which allows the entire recommendation loop to complete with no disruption to the user experience.

The three layer real-time AI architecture: a retailer’s website, Malachyte’s AI models and serving front ends, and context management infrastructure.

In addition to a fast core, a second layer of product catalog updates, inventory signals, and retailer-specific dimensional data keeps product data up to date. This flows through Cloud Pub/Sub, which offers globally accessible REST APIs that enable connections retailers can use without deep integration work. Malachyte agents run on Google Kubernetes Engine (GKE), with model inference on Google Compute Engine (GCE).

Continuous ingestion of external data, such as product catalog updates, operates through Pub/Sub’s global messaging system.

With its migration to Google Cloud’s AI architecture, Malachyte demonstrated that production AI inference and training are about more than GPUs and storage. They require real-time continuous learning infrastructure that includes a fast key-value store, a streaming layer, and a managed messaging system, all integrated with the foundation model architecture.

This approach also shows that even a small team like Malachyte’s can have a big impact in an industry. It just needs access to powerful infrastructure and core AI managed services.

Try it for yourself

Looking to shake up your industry or stay ahead of the competition like Malachyte? Try Managed Service for Apache Kafka, Cloud Pub/Sub, and Bigtable. New customers can receive $300 in Google Cloud credits.

How WPP operationalizes platform and data engineering for AI marketing

Mon, 10 Aug 2026 16:00:00 +0000

Between chaotic levels of market fragmentation and economic volatility, marketing and communications agencies can no longer rely on the human intuition they’ve traditionally used to win clients and optimize their ad spend. WPP is replacing that guesswork with an AI-powered view of shifting market dynamics, giving brands predictive certainty that lets them invest with confidence while moving at the speed of the market. That’s the value of WPP Open, its agentic marketing system.

But before it could begin applying sophisticated AI models to power those insights, WPP had to overcome a critical engineering challenge: the marketing data that made up the models was fragmented across hundreds of global agencies. While this dynamic made it nearly impossible to deploy AI tools efficiently and securely, access to models was only part of the equation. And until it built a reliable way to ingest, clean, and serve data to those models, WPP couldn’t unlock the true potential of generative AI.

To solve this, WPP partnered with Google Cloud to construct a unified data backbone and custom platform engineering path. Now, by standardizing its serverless compute patterns and data processing workflows, WPP is able to securely deploy targeted marketing campaigns in days instead of months.

Architecting a centralized, service-based data foundation

An important part of this effort was accelerating data availability and centralizing management. To do this, WPP adopted a service-based project structure for its current production environment. Rather than isolating every workload into separate silos, its engineering team centralized Google Cloud Storage (GCS) and BigQuery into dedicated, shared data projects, while also segregating the compute and processing workloads into distinct processing projects.

This structure simplified the core team’s user experience and ensured that all data consumers interacted with a unified source of truth. Because data from WPP’s various product lines lives in shared infrastructure, it was essential that security be strictly enforced at a granular level. By directly applying identity and access management (IAM) controls at the individual GCS bucket and BigQuery dataset levels, the company’s teams only see the data they’re authorized to access.

At the same time, raw data from various partners lands in dedicated GCS buckets in order to keep the raw inputs organized and isolated. From there, Managed Service for Apache Spark executes custom Apache Spark jobs to cleanse, normalize, and canonicalize information into standardized cohort definitions (SCDs). By utilizing a serverless architecture combined with Kubeflow for pipeline orchestration, WPP’s data engineering team avoided the overhead that often results from managing cluster infrastructure. This allowed them to focus entirely on the data transformation logic fueling the downstream GCS and BigQuery layers that ultimately feed the company’s audience & performance AI models.

What made our collaboration with Google Cloud successful was the balance they struck between uncompromising professionalism when it comes to best practices and timely delivery of incredibly pragmatic, real-world solutions.- Jonas DahlbaekSenior Data Engineering Lead, WPP

Standardizing data into unified cohorts

When raw data enters WPP’s processing zone, its platform converts it into SCDs that become core concepts used throughout the framework for keying purposes. These are based on five keys: age, gender, geo, product, and interest. But these underlying data definitions are fluid and continuously canonicalized to reflect evolving marketing concepts. As a result, this uniform structure allows WPP to join and aggregate data on a global scale without exposing sensitive underlying particulars or relying on shared identifiers.

The platform's core processing engine was built in type-safe Scala to ensure comprehensive visibility and compliance This custom framework tightly controls how data is transformed, and it inherently supports full source traceability while guaranteeing that every data point within the curated datasets can be traced back to its origin. This is a crucial level of traceability when building enterprise AI applications, as data scientists and auditors must understand exactly what information feeds into the models, even as WPP concurrently prepares to transition to Google Cloud Knowledge Catalog for automated, enterprise-wide data governance in the future.

Working with Google Cloud has been instrumental in accelerating and standardizing our engineering efforts. In a world where massive volumes of fragmented data present a daily challenge, having the right infrastructure is paramount to thriving in the AI age and helps our developers and AI marketers alike.
-Suleman KhanProduct Manager for OI & Google Partnerships, WPP

Standardizing the enterprise software lifecycle

For WPP, even with all these steps in place, processing data is only half the battle. To serve applications and manage the underlying infrastructure, the company’s platform engineering team developed a suite of reusable and centralized GitLab continuous integration and continuous deployment (CI/CD) templates. With this, WPP reduced the cognitive load on individual development teams and ensured that all deployments met strict corporate security standards.

These templates manage various enterprise workloads autonomously. The suite includes universal Cloud Run templates for full-stack web applications and batch data processing and scheduled pipelines. It also includes a deploy-only template for multi-stage workflows and a Cloud Run functions deployment template for event-driven microservices.

Implementing zero-rebuild promotion

Rebuilding container images in a production environment can introduce unnecessary risk and the potential for configuration drift. In order to maintain environmental consistency, WPP embraced a "build once, deploy many" methodology that applied cross-project IAM logic and Google Cloud Artifact Registry configurations.

As part of this process, developers build and test container images in the development environment. Once those exact, immutable container images are validated, they’re promote directly to production. This zero-rebuild promotion ensures total parity across deployment stages and eliminates unexpected production behaviors. The CI/CD templates also facilitate progressive traffic migration, which allowed teams to route a small percentage of traffic to new revisions before initiating a full rollout.

Immutable deployments. Traceable data. Unshakable trust. When you know exactly what goes into your AI, you can ship at the speed of light.
- Ranjith K PoldasAssociate Director , Devops (I&P), WPP Media

Automating security and intelligent networking

With this modern architecture, enterprise security acts as a foundational enabler for WPP, so it integrated Wiz security scanning directly into the pre-push phase of the CI/CD pipeline to catch vulnerabilities before code merges. The company also utilized Google Cloud Identity-Aware Proxy to enforce zero-trust access across its internal applications.

To further simplify operations, WPP adopted templates with intelligent virtual private cloud (VPC) logic. This configuration automatically identifies and resolves networking conflicts between legacy VPC connectors and modern Direct VPC access. This automated networking prevents deployment failures and accelerates the release cycle.

Monitoring operational health and driving ROI

Because a resilient platform foundation requires deep observability, WPP’s engineering team now monitors strict operational metrics instead of relying solely on deployment frequency. The team tracks request latency across p50, p95, and p99 percentiles, alongside 4xx and 5xx error rates. It also monitors container startup times to mitigate cold starts, while tracking overall CPU and memory utilization. This granularity ensures that both data pipelines and serverless infrastructure always remain highly available.

"Navigating a transformation of this scale across multiple complex workstreams—spanning data engineering, platform infrastructure, and AI integration—required more than just alignment; it demanded deep, mutual trust. Working as true partners, Google Cloud and WPP moved in lockstep to deliver production-ready platform capabilities on time."Yang Yue , Program Manager , Google Cloud

For WPP, operationalizing its data and AI stacks at this velocity provided the necessary infrastructure for its advanced workloads, and the business impact was clear and quantifiable. By building this dual foundation, the company reduced creative and strategy time from four weeks to just three hours. It also saw a 70% gain in production efficiency, a 33x increase in content volume, and a 2.8x increase in campaign return on investment. In short, by partnering with Google Cloud and implementing a broad suite of products and tools, WPP was able to quickly realize a significant ROI and boost productivity, efficiency, reliability, and security across the company.

GOL! How TelevisaUnivision streamed the FIFA World Cup to millions with Google Cloud

Fri, 07 Aug 2026 16:00:00 +0000

Live sports broadcasting represents the ultimate stress test for digital media infrastructure, where operational success or failure is measured in milliseconds and observed live by millions of viewers simultaneously. During the 2026 FIFA World Cup, the stakes reached a high for TelevisaUnivision, the leading Spanish-language media conglomerate. With Mexico serving as both a primary host nation and a core contender on home soil, fan engagement created unprecedented demand across Latin America and TelevisaUnivision's ViX streaming platform.

For a marquee broadcaster like TelevisaUnivision, high-stakes events carry direct, long-term brand and reputational implications. Audiences demand uninterrupted, pristine access to every critical moment of play. Playback interruptions during a key goal, login latency at kickoff, or degraded stream resolutions immediately impact customer satisfaction, risking subscriber churn and brand dilution. When streaming tier-1 global sports events, technical execution directly impacts consumer trust, requiring an absolute commitment to zero-downtime availability and flawless performance on the part of the provider.

Why TelevisaUnivision selected Google Cloud

Navigating Latin America's complex networking ecosystem, which is marked by heavy ISP fragmentation and cross-border transit bottlenecks, required more than a standard vendor relationship. TelevisaUnivision needed a strategic partner willing to make joint investments in network capacity, infrastructure resiliency, and custom feature development. TelevisaUnivision selected Google Cloud's Media CDN based on two foundational differentiators: its architecture, and Google Cloud’s customer focus.

Platform architecture

Media CDN provided a resilient, globally distributed infrastructure built specifically to absorb massive live-stream traffic spikes while protecting origin infrastructure. Core architectural advantages included:

In-ISP deep edge caching: Media CDN embedded cache nodes deep within local ISP networks across Mexico and Central and South America, placing video segments within a single network hop of viewers.
Direct ISP peering: By establishing direct peering connections with major regional telecommunications operators such as América Móvil and Telefônica, the architecture completely bypassed congested international transit routes.
Dedicated capacity reservations: TelevisaUnivision reserved live event capacity with dedicated allocated headroom in-region. This isolated livestream traffic from "noisy neighbor" risks and comfortably absorbed peak traffic surges.
Sub-millisecond sessions with Valkey 9.0: To handle massive traffic spikes during the World Cup, TelevisaUnivision migrated its session store to Memorystore for Valkey 9.0, operating as serverless microservices on edge compute. This architecture delivered sub-millisecond response times for critical authentication and entitlement checks, while providing automatic scaling to process peak API traffic without the need for manual capacity reservations.

Obsession for customer success

Beyond technical capabilities, TelevisaUnivision chose Google Cloud for its joint co-engineering model and deep operational alignment.

Joint 24/7 war rooms: For all 104 matches, TelevisaUnivision engineers and Google Cloud specialists operated side-by-side in unified command centers.
Proactive monitoring as a service: Google Cloud’s Customer Reliability Engineering teams provided round-the-clock proactive monitoring and automated alerting.
Joint operational authority: Combined teams performed extensive pre-tournament stress tests and simulated failovers. During live matches, unified telemetry empowered joint leads to dynamically route traffic and adjust CDN configurations instantly as regional ISP congestion emerged.

Summary

The strategic partnership between TelevisaUnivision and Google Cloud during the 2026 FIFA World Cup established a new benchmark for global sports broadcasting. Across 39 consecutive days of tournament execution, TelevisaUnivision reported that the joint infrastructure delivered:

Total matches broadcast	104 live matches
Platform availability	100% platform availability (0 downtime)
Cumulative viewership	675 million views across TelevisaUnivision & ViX

By uniting localized edge delivery, sub-millisecond serverless compute, and dedicated operational co-engineering, TelevisaUnivision and Google Cloud solidified a battle-tested blueprint for executing marquee live streaming events at record global scale.

Mirendil taps AI Hypercomputer TPUs and GPUs for pre- and post-training applications

Thu, 06 Aug 2026 13:00:00 +0000

Nearly every major AI lab uses Google Cloud infrastructure, including for training of models, inference for agents, and new frontier research. Google Cloud also continues to be the platform of choice for new, high-growth AI startups who are driving much of the industry’s research and innovation.

Today, we’re announcing that Mirendil, an exciting frontier AI lab focused on accelerating AI development, will also utilize Google Cloud’s AI Hypercomputer. This includes using a mix of Google’s TPU AI accelerators and full-stack NVIDIA AI infrastructure running on Google Cloud; this purpose-built AI infrastructure will support model pre-training and post-training applications for Mirendil.

The Mirendil team is building new AI systems that can help accelerate and democratize AI research and development. This means managing complex, end-to-end training workflows from initial model pre-training through post-training, and powering reinforcement learning on a massive scale. The ability to choose a mix of both TPU and NVIDIA’s full-stack accelerated computing platform through Google Cloud meant that Mirendil could access critical compute very quickly, and continue to match its workloads to the architecture best-suited to it over time.

We closely partnered with Mirendil on end-to-end design and deployment of combined TPU and NVIDIA AI infrastructure across compute, storage, networking, and control planes. We also collaborated on a system that uses managed training clusters running in Gemini Enterprise Agent Platform, which effectively streamlines the provisioning and management of both TPU and GPU environments for Mirendil. Mirendil is already live with a cluster of TPU v5P chips, with NVIDIA AI accelerated computing systems coming online soon.

"Progress in AI has been bounded by how fast humans can run the research loop - designing experiments, evaluating results, and iterating," said Behnam Neyshabur, cofounder and CEO of Mirendil. "We're building AI systems that can accelerate and improve that loop itself. Expanding on Google Cloud gives us the scale and flexibility to push those systems further and put frontier AI research capabilities in the hands of many more scientists and engineers to run that loop faster and at a greater scale."

You can read more about our partnership on Mirendil’s blog.

Scaling agentic AI: How UiPath built its high-performance GPU platform on AI Hypercomputer

Wed, 05 Aug 2026 16:00:00 +0000

As a market leader in enterprise agentic automation and business orchestration, UiPath is helping to pioneer an industry shift toward agentic AI. With it, the company is deploying autonomous agents to actively reason, make decisions, and execute complex business processes across its disparate systems.

This transition from simple task automation to cognitive decision-making agents requires a massive surge in computational power and powerful infrastructure that’s reliable enough for the needs of the world's largest enterprises.

Being able to orchestrate hundreds of GPUs in perfect harmony can be what makes the difference between just running a research experiment and building a global AI platform. Such orchestration requires balancing massive training jobs with real-time inference, all without letting costs spiral or latency spike.

To do so, UiPath re-architected its infrastructure to support high-scale intelligent document processing (IDP) using UiPath IXP and moved from isolated clusters to a shared Google Cloud GPU fleet, balancing A3 VM instances (NVIDIA H100 GPUs) for training with G4 VM instances (NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs) for inference. This architecture lets UiPath solve its “spiky workload” problem and count on predictable costs and open-source patterns that the company’s engineering teams can use to replicate this architecture themselves.

"Realizing the full potential of enterprise agentic AI requires an infrastructure that matches our ambition. Google Cloud provides the scale and flexibility we need to train specialized models and deploy them globally. This partnership allows us to deliver high-precision intelligent document processing and autonomous agents that don't just chat, but actively drive business outcomes for our customers." – Raghu Malpani, Chief Technology Officer, UiPath

The context: heavy-duty math

UiPath has run its full-stack automation platform on Google Cloud for years, but as its agentic AI initiatives expanded, it faced a series of new infrastructure challenges.

Core capabilities like IDP, computer vision, and LLM-powered reasoning require heavy-duty math, so UiPath’s engineering team utilizes LLAMA model grounding that allows its robots to "see" interfaces with human-like clarity. And with specialized document models built on the Qwen architecture, the team can extract valuable data from messy, real-world paperwork.

These models live on the UiPath cloud infrastructure, where cutting every possible millisecond of latency is essential. Moving from a "cool demo" to a reliable production tool without exploding costs meant the team had to rethink its underlying silicon.

The challenge: more demand than supply

In the past, when a team at UiPath needed to train a new model or run inference, it provisioned GPU nodes on demand and scaled up or down depending on whether the workloads were spiking or slowing.

This was a functional strategy when cloud capacity was cheap, abundant, and perfectly elastic. But as its AI ambitions grew, UiPath found this approach could no longer keep up with its operational complexity. It now faced three new challenges:

Spiky workloads: To ensure it had sufficient power for peak demand, UiPath often had to buy extra capacity that sat idle during quieter periods, wasting expensive headroom. The company needed intelligent, on-demand scaling that didn't require paying for silicon that wasn't crunching numbers.
Supply bottlenecks: For large-scale fine-tuning, the price-to-performance ratio on gold standard high-end A3 VM instances with 8-cluster H100s is unbeatable. But global demand for those chips has outstripped supply, making it nearly impossible to scale training efforts at the speed UiPath desired just by adding nodes.
Operational overhead: UiPath was also struggling with geographical inefficiency because stable inference demand still meant maintaining dedicated clusters in multiple regions to ensure low latency for international customers. Further, managing GPU infrastructure for both training and inference added inefficient layers of operational overhead.

The solution: a shared GPU fleet

With all of that in mind, UiPath decided to treat its GPUs as a shared strategic resource instead of a product-centric elastic infrastructure.

As a result, its engineering team designed a platform-level shared GPU fleet managed by its machine learning services (MLS) platform, which prioritizes work across teams and time windows while balancing demand across workflows. During the day, the fleet serves real-time inference and latency-sensitive workloads, and at night or during off-peak hours, it automatically switches to batch training and long-running jobs.

By coordinating workloads at the fleet level, MLS lets UiPath maximize utilization while reducing contention, all without relying on per-instance elasticity. It also enables the company to schedule capacity in advance, which improves predictability for both research and production use cases.

Why Google Cloud: AI Hypercomputer architecture

To support its growing scale, UiPath leveraged Google Cloud AI Hypercomputer, which offers a system-level approach integrating performance-optimized hardware, open software, and flexible consumption models into a unified environment. AI Hypercomputer also minimizes the friction between hardware and software layers, which allows engineering teams to focus on model performance rather than infrastructure management.

Once it settled on a shared fleet model, UiPath needed a cloud partner that could offer reliable GPU availability, competitive pricing, and burst capacity. That’s why it chose to expand its existing Google Cloud footprint with a highly specialized AI stack running on Google Kubernetes Engine .

Now, UiPath can take advantage of predictable capacity by leveraging Google Cloud’s Dynamic Workload Scheduler (DWS) to solve its supply bottleneck. The company knew Google Cloud could secure its GPU capacity consistently with notice windows measured in days. DWS allows the engineering team to schedule training runs in advance and secure capacity for short bursts, and it can now plan for capacity rather than having to react to scarcity. Today, UiPath runs all its training and most of its IDP model inference workloads on Google Cloud.

While UiPath uses A3 VM instances for heavy-duty training and fine-tuning, not all of its tasks require that level of power. That’s why it now deploys Google Cloud G4 VM instances as a net-new optimization for inference workloads. These instances offer a cost-effective balance of performance and price, which allows UiPath to run lighter inference tasks without occupying the high-performance clusters reserved for training.

"The shift to a shared fleet on Google Cloud transformed our operational model. We moved from reactive provisioning to a predictable, high-performance engine that powers our most advanced IDP and agentic AI workloads. With tools like Dynamic Workload Scheduler and a mix of A3 and G4 instances, we have the flexibility to optimize for both cost and speed. This ensures our engineers spend their time innovating rather than waiting for compute." - Arthur Wilcke, director of AI infrastructure, UiPath

Practical validation: differentiated models at scale

With consistent access to Google Cloud GPUs, UiPath can now bring advanced models into production. It can also schedule large training jobs without blocking production inference, letting it balance research experimentation with production reliability.

This allows UiPath to deliver advanced IDP capabilities that extract data from highly unstructured and variable documents with high accuracy. For example:

Omega Healthcare uses UiPath to automate over 100 million transactions with 99.5% accuracy, a 40% reduction in processing time, and 15,000 less hours of repetitive tasks per month.
Thermo Fisher Scientific uses UiPath to extract data from PDFs like invoices and purchase orders and is now able to process 53% of its invoices without human involvement, while cutting processing time by 70%.

Lessons learned

UiPath’s most significant wins so far have been increased availability and reliability. As its workloads continue to transition and it decommissions its legacy GPU resources, the company expects to see additional cost improvements.

For engineering teams looking to build similar platforms, some key takeaways include:

Decouple capacity: Instead of tying hardware to specific products, pool resources to smooth out usage spikes.
Schedule, don't react: Using tools like DWS to book compute in advance guarantees availability and stabilizes costs.
Right-size the silicon: Use A3 VM instances for training, but choose efficient options like G4 VM instances for inference.

Next steps

After its recent infrastructure evolution, UiPath is still refining its MLS platform to support the next evolution of AI innovation. To replicate this success in your own organization, use the resources below:

Build it: explore engineering patterns on GitHub.
Optimize it: Get started with Google Cloud G4 VM instances.
Learn more about UiPath - IXP

How Target is enhancing retail discovery and cutting database maintenance by 50% with Spanner Graph

Tue, 04 Aug 2026 16:00:00 +0000

In today’s retail environment, shoppers expect highly personalized product discovery experiences and conversational assistance that feels genuine, natural, and genuinely helpful. Today, successful product discovery is about understanding semantic meaning and the rich, connected relationships between products, categories, and guest intent. It is no longer just about keywords and basic browsing.

At Target, this work is handled by our Guest Product Confidence platform team. They are responsible for building the features that establish trust and guide purchasing decisions, such as ratings, reviews, and AI-driven digital shopping assistants. An exciting example of this is our Gift Finder chat agent, which we launched during the 2025 holiday season online and in the Target app to help shoppers discover the perfect items through friendly, conversational dialogue.

To deliver real-time personalization and context-rich semantic responses like these at global scale, we identified a critical architectural need to move away from a fragmented data ecosystem toward a unified data platform. We needed a solution capable of supporting high-throughput transactional workloads, highly connected graph relationships, vector similarity search, and full-text keyword search all at once.

In this post, we’ll explore how we achieved all four with Spanner.

Overcoming fragmented architecture

Previously, Target’s discovery data ecosystem relied on a combination of Elasticsearch clusters for search and inverted indexes, alongside separate NoSQL datastores for our transactional data. While functional, this fragmented architecture presented significant operational and technical challenges.

Disconnected context: Keeping separate search, vector, and transactional databases in perfect sync was a constant challenge. Siloed information led to missing context, disconnected attribute relationships, and inconsistent query results.
High operational overhead: Managing independent clusters, tuning search indexes, and handling complex, custom synchronization and aggregation logic required intensive manual intervention from our engineering teams.
Expansion bottlenecks: Expanding our retail data domains required adding new database collections, maintaining complex joins, and navigating weak transactional guarantees across our discovery and core transactional systems.
Siloed intelligence: We lacked the ability to query graph relationships, vector similarity, and keyword search indexes in a single transaction.

To build the next generation of AI-driven guest experiences, we needed to consolidate on one platform.

Building the enterprise ontology on Spanner Graph

We evaluated multiple specialized technologies, including standalone vector databases and niche graph databases. However, adding more single-purpose databases would have only worsened our operational complexity and data synchronization pipelines.

We ultimately chose Spanner Graph to build our enterprise ontology, which is a "graph-of-graphs" paradigm that allows us to construct a massive, generative AI-powered shopping graph.

By unifying our data, we bring semantic data, graph relationships, vector embeddings, and operational transactions under one roof. This establishes Spanner as our single authoritative source of truth for both transactional state and semantic intelligence.

Our high-level architecture now consists of three core pillars:

1. Enterprise augmentation
This layer captures our enterprise retail catalog, aggregates relevant metadata from multiple backend sources, and utilizes generative AI for agentic data enrichment to dramatically improve the quality and depth of the product data we ingest.

2. Unified graph, vector, and search store
Instead of shifting data across multiple databases, Spanner Graph stores our entity nodes, relationship edges, and vector embeddings in the same database engine. Spanner Graph natively supports multi-hop graph traversals, semantic vector similarity, and full-text keyword queries over our relational tables. Because this multi-model synergy is native, we get strict ACID transactions for absolute correctness across distributed workloads without the need for fragile external sync pipelines.

3. Orchestration and AI layer
This layer powers our conversational guest interfaces, utilizing rich, structured context fed directly from Spanner Graph to ground our LLMs. It extracts highly specific product relationships to power tools like the Gift Finder while governing responsible AI processes and evaluating generated outputs.

A smooth, zero-downtime incremental migration

Transitioning critical search and discovery infrastructure that millions of guests rely on required a cautious, zero-downtime approach. We executed this migration in four structured phases.

Schema and ontology mapping: We defined the specific retail entities, such as products, categories, brands, and guest preferences, and their corresponding relationships within the Spanner Graph schema.
Data integration and parallel replay: We built mutation-based data integrations in a parallel pipeline. This allowed us to continuously replay live transactional updates, apply schema transformations, generate embeddings, and write them directly into Spanner Graph in real-time.
Canary deployment: We gradually shifted live read traffic to the new Spanner Graph-backed platform, validating query performance, semantic accuracy, and database stability under real retail workloads.
Cutover and cleanup: Once performance was thoroughly verified, we fully transitioned all search and discovery traffic to Spanner and deprecated our legacy Elasticsearch stack, entirely removing the maintenance burden of those clusters.

Business impact

By building directly on Spanner Graph, we unlocked measurable technical and business outcomes:

The ultimate GraphRAG foundation: Traditional RAG relies on flat vector similarity, which often misses the structured associations between products, such as matching a toy with its compatible accessories or age-appropriateness. By combining deep graph traversals with semantic vector search in a unified GraphRAG architecture, we grounded our LLMs with highly precise context. This directly improved our recommendation relevancy, enhanced guest satisfaction, and boosted our Net Promoter Score.

Consolidated SQL + GQL interoperability: With Spanner Graph, our developers query structured relational catalog data and connected graph relationships in a single query using standard SQL and GQL (Graph Query Language). This eliminates the need for data duplication, latency, or complex ETL pipelines to bridge these paradigms.

Serverless scalability with zero growth ceiling: Spanner automatically handled massive, unpredictable traffic spikes during peak retail events like Black Friday and Cyber Monday. Spanner's built-in autoscaler dynamically adjusted computing capacity to handle burst traffic during high-intensity, limited-time promotional offers without sacrificing performance.

50% reduction in infrastructure maintenance: By consolidating our transactional NoSQL and search index databases into a single managed Google Cloud service, we eliminated the operational burden of maintaining separate database clusters. Our developers now spend 50% less time on database administration and infrastructure upkeep, allowing us to build and deploy new, customer-facing AI features much faster.

Migrating to Spanner Graph has accelerated our generative AI roadmap, serving as the ultimate proof of what is possible when you build on the right data foundation.

Want to supercharge your AI apps? It starts with databases with the right graph capabilities at virtually unlimited scale. Discover how Spanner Graph can turn data into action for your organization.

How Deutsche Bank unlocked agility with an API-ready ecosystem

Tue, 04 Aug 2026 16:00:00 +0000

When people think about digital transformation in banking, they often focus on the visible results: mobile apps and new digital services. But there's an invisible infrastructure making all these services possible: APIs. At Deutsche Bank, we recognized that APIs aren't just technical plumbing; they're the nervous system of modern banking.

A few years ago, our application landscape was dominated by monolithic systems. As we evaluated how to break them into modular, reusable APIs, one thing became clear: we couldn't just decompose our work into APIs — we needed a central API management platform (APIM) to manage what would emerge. We needed something where documentation, security policies, and governance all had to be built in from the start, not bolted on later.

The question wasn't just how to modernize, but how to best serve our customers and position ourselves for tomorrow's opportunities, especially with emerging technological paradigm shifts.

Needing a system that was adaptable, scalable, reliable, secure, and AI-ready for the demands of modern banking, we chose Google Cloud's Apigee as our APIM platform.

Building the backbone: four key capabilities

Today, Apigee manages our API ecosystem — from open banking APIs connecting us with fintech partners, to the internal microservices powering our various banking platforms, and even the client-facing applications that enable seamless digital experiences such as online banking.

Here are four important capabilities the platform offers us:

1. Unified governance without sacrificing speed

Apigee is the foundation of our API catalog. Every endpoint, version, and dependency is documented and discoverable. Development teams find and reuse existing APIs rather than rebuild functionality. We've moved from "Where's that customer data API?" — which took days — to a searchable, real-time catalog accessible to any developer.

But governance isn't about bottlenecks, it's about guardrails, and with Apigee's policy framework, we automatically enforce standards. OpenAPI specifications, schema validation, and error handling are now baked into the platform. Teams move faster because they work within consistent frameworks.

2. Security: the employee onboarding analogy

When thinking about API security, imagine onboarding a new employee. You don't give them access to every system on day one. You follow the least privilege principle, so they get exactly the permissions needed for their role. If they switch departments, their access rights will be updated. If they leave the company, access is revoked immediately. Apigee works the same way for our services and applications.

When connecting a new service — say, one that accesses customer accounts — we don't open the floodgates. Through OAuth2 scopes and API key management, we define precisely what that agent can access:

Read account balances? Yes.
Initiate wire transfers? No.
Access 90-day transaction history? Yes.
Full historical data? Only with elevated permissions.

Like employee access, these permissions are centrally managed, regularly audited, and instantly revocable. Just as we track employee activity for compliance, Apigee logs every API call to see who accessed what data, when, and why.

This becomes critical with high-volume automated systems. An automated service doesn't take breaks and can make thousands of calls per minute if misconfigured. Rate limiting and quota enforcement ensure that even when something goes wrong, the blast radius is contained.

3. Resilience and performance at scale

Banking doesn't have downtime. When customers check balances at 3 a.m. or markets surge with trading activity, our APIs must respond instantly and reliably.

Apigee's load balancing and auto-scaling evenly distribute that traffic. Health checks and circuit breakers automatically route around struggling services, and for frequently accessed data, Apigee's caching delivers sub-millisecond responses without hitting backends.

4. Observability: measuring everything

Before Apigee, understanding API performance was like assembling a jigsaw puzzle with pieces from different boxes. Now we have unified dashboards showing real-time traffic, error rates by service, usage analytics by consumer, and compliance metrics. This visibility serves operations, product managers who track partner value, and security teams who identify anomalies.

Apigee provides a central suite of capabilities for managing the full API lifecycle

The path forward

We built this infrastructure for the API economy, and in doing so, we have also built a strong foundation for the future of digital banking. As the industry evolves, this API-first approach will be critical for integrating next-generation services.

As digital banking continues to advance, a shift toward intelligent services that can react, predict, and assist in real time is underway. Capabilities such as real-time pattern recognition, predictive insights, and AI-powered assistants are becoming part of everyday digital experiences, with their visibility and impact increasing as adoption accelerates. Each of these capabilities will consume APIs — and they will introduce new requirements: ultra-low latency, high-throughput data flows, and secure orchestration across multiple APIs.

Because we invested in a flexible API platform with Apigee, we are well-positioned to adapt and optimize our infrastructure for these future needs, rather than having to rebuild it.

Emerging standards: MCP, A2A, and the future

The industry is exploring new integration standards. Protocols like Model Context Protocol (MCP) and Google's Agent2Agent (A2A) are interesting because they build on existing API infrastructure.

Our Apigee-managed APIs are well-positioned to leverage these advancements. For instance, MCP could benefit from our OpenAPI specifications, and A2A could leverage our OAuth2 framework, with both relying on the governance we've built.

We're also exploring patterns like placing new types of servers behind Apigee proxies to maintain security controls while enabling modern workflows. Our "always-API" pattern ensures that services benefit from centralized management, no matter how they are accessed.

MCP and A2A are complementary, MCP has a tools and resources focus, while A2A is focused on peer collaboration

The vision: APIs as universal interface

Every banking capability will eventually be exposed as an API. That’s not because APIs are trendy, but because they're the most flexible, composable, and governable way to share functionality, whether consumed by mobile apps, partner fintechs, analytics platforms, or other automated agents.

At Deutsche Bank, this shift is already taking shape. The same API foundation that powers our core platforms is now enabling our evolution toward more intelligent, AI-supported services across the bank. That foundation provides the consistency, governance, and scalability needed to bring these capabilities to life, ensuring that as new intelligent services emerge, they can be integrated seamlessly, securely, and at enterprise scale.

Apigee makes this possible by providing governance that scales across all use cases. It's not about controlling innovation; it's about enabling it safely.

Lessons learned

Invest in excellent documentation. Semantic summaries and clear schemas aren't extras; they're foundational for both developers and AI.
Treat security like employee onboarding. Least privilege and role-based access apply equally to APIs.
Observability is a competitive advantage. Unified analytics enable data-driven decisions.
Plan for the future now. Your API management infrastructure becomes your advanced integration layer.
Stay curious. Experiment with emerging standards. Flexibility wins.

Conclusion

We're at an inflection point. The API economy enabled fintech and open banking. Now, the same infrastructure can serve as the backbone for the next wave of innovation. Our investment in the API platform wasn't just about managing APIs better; it was about building a foundation for whatever comes next.

As the industry transforms, we’re ready. The future belongs to organizations that move fast without breaking things. For us, that future is powered by Apigee.

Best Buy scales AI workloads and secures access with Workforce Identity Federation

Tue, 28 Jul 2026 16:00:00 +0000

As Best Buy expanded its use of Google Cloud for advanced analytics and AI, its technology teams faced two significant scaling challenges: Mitigating risk and managing administrative friction when syncing thousands of backend users from Microsoft Entra ID.

The retailer solved both problems and paved the way for a massive cloud expansion by implementing Google Cloud's Workforce Identity Federation. This direct approach allowed developers to access cloud resources securely using their existing Microsoft credentials without a separate identity store, giving technical leadership confidence that access remains strictly controlled, auditable, and manageable at scale.

Replacing service accounts with direct federation

Best Buy historically maintained complex synchronization pipelines to copy backend users from Entra ID to Google Cloud. Because the organization used Cloud Identity without a Google Workspace deployment, it needed a more direct approach. Previously, Best Buy's Power BI integration with BigQuery relied on service account credentials.

This pattern can work at a small scale, but quietly becomes a liability as your team grows.

Manually rotating keys for service accounts meant tracking the credentials each team held, and accepting that every key was a potential security vulnerability. Service account keys created daily friction for the Best Buy security and platform teams, and the technical debt compounded as data access requirements grew more complex.

To support tens of thousands of users, Best Buy modernized its identity architecture. The team adopted Workforce Identity Federation to federate existing Entra ID identities directly into Google Cloud.

Now, when developers access BigQuery through Power BI, they authenticate as themselves using their existing Entra ID identity. They no longer need to rotate keys, worry about credentials exposed in chat messages, or guess who performed an action in the audit log.

The architecture relies on two components working together: Entra ID handles authentication, Workforce Identity Federation brokers the trust relationship between Entra ID and Google Cloud. This federation is stateless on Google's side. It validates tokens at the moment of access instead of syncing user records. Removing the service account key layer greatly reduces the credential management burden.

Architecture

The diagram below shows how identity flows from Entra ID through the Workforce Identity Federation to the services teams use at Best Buy. The key change from the previous approach is the removal of the service account key layer entirely; there is no credential to manage between Entra ID and Google Cloud.

Identity flows from Entra ID through the Workforce Identity Federation to the services teams use at Best Buy

Key implementation decisions

When implementing this architecture, Best Buy made several important technical choices:

Separate provisioning and SSO apps in Entra ID: The configuration follows the Entra ID provisioning and single sign-on (SSO) setup guide. You should separate the provisioning application from the SSO application in Entra ID. Running them as two distinct enterprise apps provides a cleaner separation of concerns; provisioning changes do not affect SSO configuration, and vice versa.
Place the automation OU carefully: You need to place the Entra ID provisioning service account in a separate organizational unit (OU) and explicitly disable SSO for that OU. This prevents a bootstrapping problem: If you enforce SSO globally, the provisioning account cannot authenticate to set up the provisioning in the first place.
Understand that syncless means stateless on Google's side: Workforce Identity Federation does not create or maintain user records in Cloud Identity. It validates tokens at the moment of access. This makes the architecture viable for Best Buy's target scale, because it eliminates synchronization lag, stale record cleanup, and separate provisioning pipelines.

Secure authentication for developers

For developers, the change was practically invisible. They authenticate once through their corporate Entra ID credentials, and access to BigQuery works automatically, whether through Power BI or direct API calls. The SSO experience matches everything else they access through their Microsoft identity.

For the security and platform teams, the benefits are significant. The attack surface from credential management disappears. Audit logs now show individual users instead of shared service account identities, and you can revoke access quickly based on the enterprise identity lifecycle rather than waiting for manual key rotation.

If you currently manage service account keys for developer access to Google Cloud, moving to Workforce Identity Federation is worth the effort. You gain significant security benefits, and the operational simplicity grows as your team expands. Best Buy is currently scaling this secure access to a broader workforce to power its future retail operations.

Expanding Workforce Identity Federation support

Google Cloud continues to make it easier for all organizations to bring their own identity providers. Recent updates simplify the setup for Ping Identity users and extend access to online billing accounts.

Ping Identity integration: If you use Ping Identity, you can follow a new, dedicated setup guide to configure federation. This guide provides step-by-step instructions so you can securely connect your workforce to Google Cloud resources.
Online billing support: Google Cloud now supports customers with online billing accounts. You can use Workforce Identity Federation for secure, syncless access without needing an enterprise billing agreement.

Get started

Google Cloud is committed to removing friction from cloud adoption and making it simpler for organizations to secure their environments. To explore these new capabilities and connect your organization's identity provider, read more about how Workforce Identity Federation allows you to federate identities directly, and explore our supported Google Cloud services.

The Blueprint: How Voicify makes AI-enabled ordering a delight for customers

Thu, 23 Jul 2026 16:00:00 +0000

Welcome to The Blueprint, a new feature where we highlight how Google Cloud customers are tackling unique and common challenges across industries using the latest AI and cloud technologies. We hope to inspire others looking to innovate in their work.

Founded in 2018, Voicify reimagines the traditional phone call with the goal of transforming every call into a seamless and engaging experience.

The challenge:

When we started Voicify in 2018, our vision was to help organizations build confident, pragmatic, and technically grounded voice-driven assistants for any channel, including phones and chat. But the pandemic changed everything.

We shifted our focus to telephone use cases primarily in the restaurant and healthcare sectors where, at the time, call volume and staffing posed significant challenges. Restaurants could potentially miss up to 20% of their calls and lose orders as a result, and healthcare providers struggled to keep up with call volume with the required 100% accuracy when integrating appointment information into a practice management system. We realized that specialized, purpose-driven AI assistants were the key to businesses maintaining excellent service at scale.

To succeed, we had to overcome four primary challenges:

Transactional precision: Our voice assistant needed to reason with complex customer requests against point of sale and practice management systems with 100% accuracy.
Traffic spike management: Our LLM usage needs to be provisioned accurately to keep costs down and maintain customer services in spite of the common (and extreme) spikes in traffic seen in restaurants and healthcare organizations.
Latency: Any delay in the assistant’s response can cause customers to hang up. We needed superfast time to first token, with minimal delay from when a user sends a voice or text request to when the AI model generates its first piece of output.
Security and compliance: Since our founding in 2018, we’ve ensured that we’re HIPAA, SOC2, ISO27001, and PCI-compliant, and that our security is enterprise-grade. We needed architecture and infrastructure that employs all possible safeguards to safeguard data integrity and security.

The solution:

Our conversational orchestration platform builds and validates restaurant orders against a point-of-sale system before submission to ensure accuracy. Under the hood, Gemini Flash, served via Gemini Enterprise Agent Platform, vastly improves latency, minimizing user wait times and preventing hang-ups. With it, we also see approximately 25% to 30% savings compared to our previous use of other LLMs, and with greater reliability too.

To grow the business — and call volume — and to handle traffic spikes, we switched from Google AI Studio to Vertex AI and its current incarnation in Gemini Enterprise. We wanted the enterprise guarantees the latter provided, which we needed for scaling as well as for security and compliance for our healthcare clients. Specific Gemini Enterprise Agent Platform features help us manage high call volumes without experiencing service interruption or dropped responses.

These enterprise-grade services may have carried an increased cost over AI Studio, but they were well worth it to ensure reliable uptime, and the premium pay-as-you-go feature made scaling much easier for us. For example, we used a combination of provisioned throughput and premium pay-as-you-go with Vertex AI to accommodate all-time high usage the day before Thanksgiving, and we saw no rate limiting issues.

The architecture:

The outcome:

Gemini has reduced the burden on our in-house programmatic tools for pulling context and building menus. We’ve seen great improvements in performance and reliability, with lower latency and greater reliability with Gemini. And, the increased stability of our Gemini-powered assistants has made client onboarding much more efficient. Now it only takes one to two days to get a restaurant ready to test after gaining access to the POS system, down from what previously took one to two weeks.

With Google solutions for scale and enterprise-grade service, we’ve optimized our critical time-to-first-token metric, minimizing customer wait times. Using Vertex AI’s provisioned throughput and pay-as-you-go features, we’ve ensured 100% uptime, prevented dropped responses and rate-limiting issues, even during periods of all-time high usage. We’re now able to easily manage the spiky nature of restaurant traffic.

In terms of technology, we anticipate moving beyond conversational order capture to more proactive assistance, using context from conversations or POS activities. Your typical Friday night order from your favorite Japanese restaurant? Someday soon it might be Voicify’s voice assistant proactively placing it for you.

The details:

Our industry focus presents a few unique challenges that we had to spend time solving within the backend.

The core component of our Voicify solutions is our voice orchestration platform, which manages the entire phone AI stack and is designed for enterprise-grade scalability and security. This is also the node where industry solutions are called depending on user needs.Our voice orchestration platform sits close to the customer and coordinates backend services like Gemini and the different components of the voice assistant. We use it to manage functions like automated speech recognition, text-to-speech, and text generation, which is not purely generative but includes programmatic elements.

One of the unique architectural decisions we made was how to manage large, complex restaurant menus. We decided to avoid putting the entire menu into a single prompt, and we include only certain information in the initial prompt and then gather more details as the conversation progresses. This improves response times and helps manage the complexity of larger orders by focusing on only the relevant parts of each menu in a given interaction.

We also designed the architecture from the outset of our company to meet the high standards of enterprise clients for security and compliance, particularly in healthcare. We are making sure that our scalability is enterprise-grade. Architecturally we’re also employing all safeguards to ensure data integrity and safety too. Lastly, our platform is designed to support a multicloud environment as part of our strategy for achieving the highest possible level of availability.

From maintenance to innovation: Checkout's migration to Managed Service for Apache Airflow

Wed, 22 Jul 2026 14:00:00 +0000

Data engineering teams often face a “Day 2” operational reality after building a data platform: the ongoing work of maintaining the orchestrator itself.

For the Data Platform team at Checkout.com, managing a self-hosted Apache Airflow environment on another hyperscaler was consuming time the team wanted to spend elsewhere as server management, patching, and incident response were pulling focus from building pipelines.

By migrating to Managed Service for Apache Airflow (Gen 3), Google Cloud’s fully managed Airflow service, Checkout.com transformed its reliability and cost structure. Here’s how they built a more scalable, cost-efficient, and robust data foundation.

The starting point: self-managed Airflow

Before the migration, Checkout.com ran Airflow on self-managed infrastructure. While functional, maintaining the underlying resources required significant attention. Patching, upgrades, and server management created regular interruptions.

Operational data from the past year illustrates some of the challenges the company was navigating:

Reducing operational friction: In its self-managed environment, Checkout.com faced stability challenges, particularly during high-load periods.
Complex dependency management: Upgrading packages and ensuring compatibility was a constant, manual struggle. With Managed Airflow (Gen 3), the company was able to simplify this by handling dependencies at the image level, ensuring seamless compatibility out-of-the-box during routine environment upgrades.
DAG sync time: Syncing DAGs to the scheduler took approximately six minutes after deployment to S3, which affected iteration speed.
Manual processes: Scaling required manual intervention, and onboarding new teams meant manually creating secrets and variables for dbt.

The solution: Managed Service for Apache Airflow (Gen 3)

Checkout.com’s team migrated to Managed Airflow to offload infrastructure responsibility and take advantage of Google Cloud's managed scalability. The results were immediate and measurable across three areas: reliability, cost, and developer velocity.

Dynamic scaling in action

In the previous elastic container service setup, the team allocated the maximum number of workers required for peak loads. This meant paying for peak capacity around the clock, regardless of actual usage.

Managed Airflow provides built-in dynamic scaling, eliminating the need for manual resource management. The environment automatically adjusts the number of workers based specifically on the workload demands. When tasks spike, the system scales up; when they drop, it scales down to save resources. Similarly, moving from fixed provisioning to dynamic scaling reduced monthly costs by an estimated 30%.

Reliability and DAG isolation

Achieving increased stability was a primary driver for Checkout.com’s migration since in the past, a single problematic DAG could affect its entire environment.

Managed Airflow introduced a number of critical architecture improvements:

DAG isolation: Each DAG runs in its own execution environment. If one DAG fails or consumes excessive resources, it doesn’t affect the entire environment.
Managed operations: Google Cloud handles patching and upgrades during scheduled windows, removing the need for manual upgrade management.
Improved visibility: Integration with Google Cloud Monitoring and Cloud Logging provides clear visibility into task execution. Engineers can now debug issues independently without escalating to the platform team.

Faster developer workflows

The migration also improved day-to-day workflows for Checkout.com’s data engineers.

Faster deployments: Using Cloud Storage for DAGs enabled near-instant syncing.
Simpler onboarding: Teams no longer needed platform support to create variables before onboarding.
Modernizing dbt execution: One of the company’s most significant wins was changing how it runs dbt. Previously, its engineers had to manually install and manage complex virtual environments for every supported dbt version. By leveraging containerized dbt runs, Managed Airflow (Gen 3) eliminates dependency bottlenecks. This ensures complete dependency isolation, allowing teams to run any required dbt model with minimal setup and no manual infrastructure overhead.
Environment updates: The company no longer needs to redeploy the entire Airflow environment to add new roles or update Python packages.
AI-powered troubleshooting with Gemini Cloud Assist: In a self-managed environment, a failed task often triggered a frantic hunt through fragmented logs and metrics. With Managed Airflow, Checkout.com can initiate a Gemini investigation directly from its Airflow DAG UI in the Google Cloud console.

Gemini doesn't just provide generic error messages; it generates a scorecard that evaluates different hypotheses with both supporting and contradictory evidence, which can drastically reduce mean time to recovery.

Conclusion

For Checkout.com, the move to Managed Airflow (Gen 3) marked a strategic shift, one that freed its engineers to focus on delivering value.

"With Managed Service for Apache Airflow, we’ve achieved significant improvements in efficiency, scalability, and reliability. Managed infrastructure, automated scaling, faster deployments, and isolated execution environments have transformed how we operate." — Keisi Mancellari, Data Platform Engineer, Checkout.com

With a stable, scalable, and cost-efficient platform in place, Checkout.com is now able to focus on the future of its data pipelines, confident that its orchestration layer is ready for whatever comes next.

Learn more about Managed Service for Apache Airflow and how it can support your data platform.

_{Special thanks to the following contributors to this post: Serge Bouschet and Keisi Mancellari}

SOCRadar powers rapid threat detection with AlloyDB and Gemini Enterprise

Wed, 01 Jul 2026 19:00:00 +0000

Editor’s note: SOCRadar is a leading cybersecurity company that provides threat intelligence to businesses worldwide. As the volume of cyber threats continued to grow, SOCRadar needed to modernize its data infrastructure to deliver faster insights to its customers. By migrating from PostgreSQL to AlloyDB, SOCRadar achieved a 20x performance boost, reduced operational overhead, and is now better positioned to innovate and grow.

How SOCRadar supercharges rapid threat detection with AlloyDB

SOCRadar provides external threat intelligence to help organizations across 30+ countries defend against cyberattacks. On the front lines of cybersecurity, timely intelligence is everything and a delay of a few minutes can mean the difference between a blocked exploit and a full-scale breach.

As SOCRadar’s business scaled and cyber threat volumes exploded, their on-premises, self-managed PostgreSQL database hit a wall. The database simply couldn't keep pace with the simultaneous demands of high-velocity data ingestion and heavy, real-time analytical queries. This created a severe data bottleneck, slowing down the delivery of critical insights to customers and pulling engineers away from innovation to focus on constant manual database tuning.

Evaluating database alternatives: The hunt for scalability

The engineering team realized their traditional PostgreSQL environment had reached its absolute performance limits. To scale, SOCRadar needed a high-performance fully managed database that could dramatically slash operational overhead while elegantly handling a complex, hybrid workload.

They evaluated alternatives and selected Google Cloud's AlloyDB for PostgreSQL. Because AlloyDB is fully PostgreSQL-compatible, it offered a low-risk migration path while promising a specialized architecture built to handle both high-volume transactions and real-time analytics simultaneously. To accelerate the transition, SOCRadar partnered with NGC, a Premier Business Partner, who meticulously validated the architecture before executing a precision cutover with minimal downtime.

Taming a "triple-threat" workload

Migrating to AlloyDB transformed how SOCRadar processes massive, diverse cyber telemetry. Today, AlloyDB effortlessly manages what SOCRadar’s engineering team calls a "triple-threat" query environment, maintaining sub-second lookup latency even as processing volumes scale.

To understand the performance leaps, it helps to separate the system’s velocity (handling live data streams) from its depth (analyzing historical data):

High-Velocity Transactional Ingestion (OLTP): The platform constantly ingests real-time telemetry from thousands of disparate, fast-moving sources—including Dark Web forums, botnet logs, and social media feeds. AlloyDB handles these continuous INSERT and UPSERT operations with a 3.2x boost in live ingestion velocity, ensuring that the newest threat indicators are immediately recorded and available for detection.
Real-Time Operational Point-Reads: When a security analyst is actively investigating a live incident, speed is everything. Baseline performance testing under zero-load conditions for random ID lookups on indexed fields (e.g., querying a specific Indicator of Compromise by ID) showed that standard queries requiring 3 to 3.5 seconds were completed in just 1 second on AlloyDB.
Deep Analytical Aggregations (OLAP): When a client requests a complex sectoral report such as correlating the most prevalent attack vectors in the finance sector over an entire year, the database must execute deep scans across vast historical datasets. Leveraging AlloyDB’s built-in In-Memory Columnar Engine, these analytical queries run up to 20x faster than standard PostgreSQL.

More than just speed: Reclaiming 45 TB and 75% of DBA time

While the raw performance gains were massive, the operational and financial impact completely changed how SOCRadar's engineering team works day-to-day.

Thanks to AlloyDB's advanced automation, including intelligent memory management and write-ahead log (WAL) optimization, the need for constant, manual database tuning evaporated. The database administrator's (DBA) workload dropped significantly, requiring a system health check just “about once every two or three days." This freed up 75% of SOCRadar’s DBA resources, allowing them to pivot away from maintenance and focus entirely on core platform innovation.

Financially, AlloyDB’s dynamic storage management solved a massive cost efficiency issue. Unlike traditional database environments that lock you into paying for fixed, provisioned storage even after data is purged, AlloyDB automatically scales storage down to match actual data footprints. By clearing out legacy, unnecessary logs, SOCRadar was able to instantly reclaim over 45 TB of storage, achieving massive, automated cost optimization.

Fighting alert fatigue with integrated Gemini Enterprise Agent Platform

Beyond scaling infrastructure, AlloyDB has allowed SOCRadar to redefine the core architecture of their threat response using artificial intelligence.

Security operations centers (SOCs) globally are plagued by "alert fatigue"—the sheer volume of security alarms makes it easy to miss a critical attack. To solve this, SOCRadar integrated Gemini Enterprise Agent Platform as a core component of their solution architecture, linking it directly to their Alarm Management framework running on AlloyDB.

By running Gemini AI-native filtering directly on their active data workloads, SOCRadar can automatically distinguish between true positives and benign false alarms. The AI categorizes, filters, and routes alerts before they ever reach the end-user. This ensures security analysts are insulated from noise and receive only the most critical, validated, and actionable intelligence.

Expanding capabilities: The future of agentic threat hunting

With a high-performance foundation firmly established, SOCRadar’s dedicated AI team is transitioning from passive analytics to active automation. The company is currently testing Agentic AI workloads, with plans to roll them into production in subsequent phases.

By integrating Real-time Data Agents with Gemini Enterprise and AlloyDB, SOCRadar is transforming with autonomous agents that don't just store data, but actively hunt threats, reason over context, and take action. Their upcoming production roadmap includes:

Natural Language Querying (NLQ): Allowing analysts to conduct rapid threat hunting using conversational language, lowering the technical barrier to querying massive database sets.
Intelligent Semantic Similarity Search: Leveraging native vector embeddings and Gemini Enterprise to allow Data Agents to independently surface hidden patterns across historical logs that traditional keyword searches would miss.
Automated Incident Summarization: Instantly transforming hundreds of lines of complex, deeply technical logs into concise, plain-language executive summaries for security analysts during critical incidents.

By consolidating transactional velocity, historical depth, and built-in AI intelligence into a unified platform, SOCRadar has eliminated its data bottlenecks and built a highly automated, future-proof framework for global cybersecurity defense.

Ready to modernize your database infrastructure? AlloyDB provides a fully managed, PostgreSQL-compatible database with high performance for transactional, analytical, and AI workloads. Learn how you can reduce costs, eliminate management overhead, and build intelligent applications.

How Schrödinger sped up molecular discovery by 4x with Alphaevolve

Tue, 30 Jun 2026 16:00:00 +0000

Computational chemistry researchers have traditionally faced a frustrating trade-off when simulating molecular interactions: use fast classical force fields that sacrifice precision or rely on accurate quantum-mechanical methods that run too slowly on large jobs.

Machine-learned force fields (MLFFs) close that gap by training neural networks on high-fidelity quantum data. When it comes to modern drug discovery and materials design, though, there’s demand for even faster processing speeds to handle massive chemical libraries involved. To overcome such performance constraints, Schrödinger partnered with Google Cloud to deploy AlphaEvolve, an evolutionary AI coding agent developed by Google DeepMind that iteratively generates and refines algorithms to find the most efficient code path overcoming the algorithmic bottleneck.

A collaborative duet with AlphaEvolve

Schrödinger — a leader in developing scientific software for over three decades — identified two critical algorithms within their MLFF training pipeline that limited performance: neighbor list computation and Ewald summation. These algorithms aggregate data from atomic neighbors and calculate long-range potentials, but both became limiting factors in training and inference speed.

Schrödinger's primary technical goal was speeding up AI model training for energy and force calculations. Specifically, they targeted the Ewald summation, a critical but computationally demanding function used in molecular mechanics. The Ewald sum was the main performance constraint in Schrödinger's PyTorch code. It had no established vectorized algorithm and often relied on simple for-loops that ran slowly on large simulations.

By incorporating AlphaEvolve into their models, the system could generate a batched implementation of the Ewald summation using parallel batch matrix multiplication. This would evolve the PyTorch code to outperform existing custom kernels.

Evaluation metrics

Schrödinger used a rigorous multi-layered evaluation framework to confirm the evolved code was both performant and scientifically accurate:

Inverse time (primary metric): The core objective was to maximize throughput by reducing calculation time, from a baseline score of 7.9.
Functional correctness: All evolved programs had to pass a full test suite, including regression tests on complex systems such as disordered water models.
Success rate: This was measured by the share of programs that were both functionally correct and faster than the baseline.

“AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months.” — Gabriel Marques, technical lead of machine learning, Schrödinger

Results: a 4x speedup and breaking bottlenecks

By applying AlphaEvolve, Schrödinger replaced simple for-loops in the Ewald summation code with parallel batch matrix multiplication. This optimization raised the program success rate from less than 1% (40 out of 5,000 evaluations) to more than 60%, while improving the performance metric from the baseline of 7.9 to nearly 30.

Optimizing these foundational algorithms delivered a 4x speedup in both MLFF training and inference. This acceleration lets researchers compress molecular screening timelines and directly benefits several key research areas:

Drug discovery: Identifying viable therapeutic candidates quickly to address urgent medical needs.
Catalyst design: Developing efficient chemical processes for industrial applications.
Materials development: Designing next-generation materials with custom properties for electronics and energy storage.

The next evolution

Schrödinger plans to apply this evolutionary approach to custom GPU kernels to test whether AI-generated code can outperform human-engineered implementations.

Read the full technical paper on AlphaEvolve to learn how evolutionary AI agents optimize scientific codebases, or contact the Google Cloud AI team to discuss accelerating your research workflows.

How growing UK midsize businesses are building in the AI era

Wed, 17 Jun 2026 08:00:00 +0000

The UK’s 5-million-plus small and midsize businesses and enterprises (SMBs) are the backbone of our economy. Today, we’re seeing these critical businesses begin to put AI to work, to operate more efficiently, move faster, and ultimately deliver better outcomes for their customers.

This shift is driven by tangible day-to-day results. According to recent research from Enterprise Nation published in partnership with Google, 71% of AI adopters surveyed in the UK say the technology helps them save time on routine tasks, and 64% report a direct boost in productivity. On top of this, AI-enabled productivity tools (like Google Workspace with Gemini) are delivering a 20% boost in productivity for SMBs, which effectively hands them back one full working day every single week.

At Google Cloud, we have a front row seat to this shift: SMBs have long utilized platforms like Google Workspace, and today they’re transforming with Google’s AI platform and models. In fact, we’ve seen the number of UK-based SMBs using Google Cloud AI nearly double year-over-year. This includes our Gemini models and products like Gemini Enterprise and AI Studio, which are helping SMBs do things like:

Roll out better customer support systems to help escalate and resolve customer support calls more quickly.
Automate repetitive actions in areas like payroll and accounting.
Help more employees understand and leverage data at work — even those not trained as data analysts.
Rapidly create and implement new designs for marketing collateral.
Help more people build their own AI agents to help them in their everyday jobs.
Conduct complex research projects at a speed and price point previously unavailable.

At today’s Google Cloud London Summit, we’re showcasing a number of innovative SMB customers who are actively using our AI tools to transform how they work, including companies who have recently expanded their work with us:

Neural Alpha, a sustainability fintech company, is using Gemini models to read unstructured environmental and corporate sustainability reports to automatically find and organize thousands of key facts, cutting months of slow, manual research down to a fraction of the time.
Sep 2, a digital security provider, uses Gemini Enterprise to deploy autonomous AI agents for 24/7 threat monitoring — accelerating incident detection and quickly neutralizing security threats for its customers.
Sunhouse, a strategic brand design agency, uses Gemini Enterprise to easily find archived design work stored on Google Drive, enabling its teams to spend less time hunting for files and more time growing its business with global brands.
Terrapinn, a global B2B events company, is transforming its operations by leveraging Gemini models, NotebookLM, Looker, and BigQuery to turn manual tasks into automated workflows, accelerating how its teams design, market, and deliver world-class conferences.
VoCoVo, a telecommunications provider, is integrating Google Cloud AI across its systems to turn isolated data into actionable intelligence and build autonomous workflows, streamlining routine operations so their team can focus on high-impact innovation.

Empowering Your Team: AI Upskilling Resources for Growing British Businesses

To help midsize teams maximize their impact and confidently navigate the modern AI landscape, we’ve developed a suite of dedicated, no-cost upskilling resources. Whether you want to train your existing teams or democratize data tools across your entire workforce, these programs will help you build an AI-ready organization:

SMB-Focused Programs: Explore our new SMB Learning Path or enroll in the Gemini Enterprise Agent Ready (GEAR) program for specialized training in agentic AI.
Google Skills for Organizations: Access our no-cost, on-demand learning platform featuring over 3,000 AI courses and hands-on labs created by experts at Google Cloud and Google DeepMind.
Get Certified: Ready to validate your team's expertise? This premium, cohort-based program offers instructor-led training, technical mentorship, and AI-infused skill badges designed to prepare your team for industry-recognized certifications.

By offering a full suite of SMB technology and training — from productivity in Workspace, to all our Ads services, and now powerful AI tools — Google is helping small and midsize firms thrive, no matter where the future takes us.

From AI potential to agentic reality: Driving the UK’s next chapter

Wed, 17 Jun 2026 08:00:00 +0000

The United Kingdom, and London in particular, continues to be one of the great hubs for AI development in Europe and the world. We’re home to Google DeepMind, of course, as well as significant AI unicorns — and Google Cloud customers — like Ineffable Intelligence, which is today announcing an important partnership with us.

A year ago, we joined you for the London Summit to showcase the vast potential of generative AI, including a major investment in upskilling the UK civil service. Today, as we welcome our partners once again to the historic vaults of Tobacco Dock, that potential has become an industrial-scale reality. In my conversations with leaders across both Whitehall and The City, the focus has moved from chatbots and media experiments to full-production execution. This is the moment of the agentic enterprise, where we shift from systems that simply chat with us to systems that can reason, plan, and execute multi-step workflows.

This transition is the cornerstone of the UK’s projected £400 billion economic boost from AI by 2030. At Google Cloud, we are the only provider offering the full integrated stack — custom silicon, frontier models, and planet-scale infrastructure — required to turn the Agentic Enterprise into a reality.

The new frontier of British enterprise and research

The banking sector is a key proving ground for this shift. And HSBC, one of the largest and most important financial institutions in the world, is showing the way. Today, we’re announcing a multi-year transformational partnership with HSBC to accelerate AI adoption across HSBC’s products and services globally. This new collaboration will further accelerate the shift towards AI-enabled ways of working across HSBC’s global operations.

HSBC will work with Google Cloud and Google DeepMind engineering teams to collaborate on new AI-powered tools and programmes, with access to Google’s latest agentic AI capabilities – including Gemini models and the Gemini Enterprise Agent Platform. The initial delivery focus on three areas: hyper‑personalised wealth management support, stronger financial crime risk management, and AI tools to enhance frontline/relationship manager client service

UK startups also continue to break new ground with technology, and AI in particular, as demonstrated by the work of frontier labs like Ineffable Intelligence. The company, which launched earlier this year, has chosen Google Cloud as its preferred cloud partner, utilizing Google’s full stack of AI-optimized hardware and tools to build and train Ineffable’s first generation of foundational models.

Led by David Silver, a former Google DeepMind researcher who was instrumental in the AlphaGo project, Ineffable Intelligence is taking a unique approach to AI development. The team are building systems that learn primarily through their own experience through reinforcement learning, instead of relying on the large-scale human-generated datasets behind language models. The ambition is to create a “superlearner” that develops knowledge through trial and error. This year, Ineffable Intelligence set a record for a European seed funding round of $1.1 billion, and now Ineffable Intelligence will support its training work by deploying one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 platform on Google Cloud, delivering massive computational scale.

To move from experimentation to true industrial production, businesses need more than just models; they need a roadmap. To help show them the way, we’re expanding our partnership with Deloitte, which will open a new AI Studio at its London campus. Developed in collaboration with Google Cloud, the studio will help British organisations move beyond AI experimentation to deploy autonomous, action-oriented AI systems at scale.

Deloitte is also committing to upskill 1,000 members of its UK AI and data workforce on Gemini Enterprise. This certification program will ensure that Deloitte’s AI and data engineers’ are equipped with the technical expertise to implement Google’s most advanced agentic architecture, providing UK clients with one of the largest pools of certified AI talent in the region.

Building a future-ready public sector

The blueprint for a modern digital government requires moving away from rigid legacy contracts toward agile, AI-driven public services. In collaboration with the Ministry of Housing, Communities and Local Government (MHCLG), the i.AI incubator, Google Deepmind, and Faculty, we are delivering tangible public sector reform and tools for reinvention that directly support the national goal to "get Britain building."

Agencies like MHCLG are already using a tool called Extract which was built using Google technology to help transform planning processes by reducing document processing times from two hours to just two minutes. Simultaneously, we are supporting trials of an AI planning tool — co-created with local planning authorities in Barnet, Dorset, and Camden — which aims to cut decision times for everyday applications by 50%. Furthermore, the Department for Transport (DfT) is utilizing Gemini to streamline public consultation analysis, a move projected to save £4 million annually.

Innovation on this scale also requires a secure, sovereign foundation. That is why Google Cloud is working to strengthen our UK data residency commitments, including measures like making Gemini 3.5 Flash, which features in-country AI processing, available by late June 2026 for sensitive sovereign use cases. We are giving British organizations the confidence to innovate within strict compliance boundaries.

To help keep businesses safe from the challenges posed by bad actors using AI and other digital threats, we also recently announced a comprehensive AI-powered cybersecurity platform — Google AI Threat Defense — which combines Wiz, Mandiant, Gemini & CodeMender to find, fix, and protect our customers from vulnerabilities.

Proven impact from the high street to public service

Autonomous agents are no longer a future prospect; they are delivering value across the UK economy today. Our work with THG Ingenuity, an ecommerce solutions provider, has delivered an 8x higher conversion rate via its AI Shopping Assistant. Starling is similarly empowering customers with "spending intelligence" tools for instant habit analysis around purchases and expenses. And Rightmove, has launched a beta version of an AI-powered conversational property search, built with Google’s Gemini models, enabling users to search for homes in their own words.

The breadth of this impact is visible across every sector: Kingfisher is pioneering retail-specific agentic applications; Openreach is driving field service optimization in telecommunications; andUnilever is using AI at scale across the entire value chain to drive growth and build desirable brands in the new era of consumer goods.

Meanwhile, VMO2 is streamlining complex data operations; Vodafone is executing a $1 billion partnership to redefine network performance; and WPP is integrating Gemini across creative workflows, whether that's generating high-fidelity campaign assets at speed and scale, powering AI agents, or training robotic camera operators.

Empowering the engine of growth for small to medium businesses and startups

The true measure of Britain’s AI success lies in its small and medium enterprises and startup ecosystem. Our AI Works research highlights a pivotal moment: AI has the potential to boost productivity for small and medium enterprises by 20% and unlock £198 billion in output for the UK economy. With 56% of smaller firms already seeking guidance, we have launched the AI Works for Britain upskilling initiative to ensure no business is left behind.

We also continue to foster the next generation of British unicorn startups through our ongoing partnership with Tech Nation at the London AI Hub. This sustained commitment ensures founders have the resources and community needed to scale, and this September, we will further this mission by hosting the Gemini Startup Forum: Cybersecurity in London to help startups build secure-by-design AI applications.

The Model Garden at Platform 37

Our belief in the UK’s potential is reflected in our physical footprint, too. We are continuing to invest in the UK's digital infrastructure to support growing demand: Our state-of-the-art data center in Waltham Cross launched in September 2025, a key part of our two-year, £5 billion investment to help power the UK's AI economy. And earlier this year, we opened our new office in London in Kings Cross, Platform 37, along with plans for The AI Exchange, a new public space dedicated to deepening understanding of AI.

Building on this momentum, we are excited to introduce The Model Garden at Platform 37, launching in the fourth quarter of 2026. This London-based hub is far more than a physical space; it serves as a strategic investment designed to fundamentally elevate how we engage with our most important customers. Blending the timeless aesthetics of a classic English garden with immersive, high-tech innovation — from living digital walls to a three-story atrium — The Model Garden acts as a physical marketplace for our best ideas.

The blueprint for the agentic enterprise

For UK businesses, civic leaders, and organizations to continue to lead in the AI moment, they must not only rethink the technology they use but also fundamental aspects of how we work. As we support thousands of organizations and millions of teams here and around the globe, we see three core strategies helping achieve success with AI:

Culture: We must reimagine our organizations for the future. True transformation means getting teams excited, enabled, and equipped to work with AI agents in completely new ways. It is about human-AI collaboration, not just automation.
Responsibility: We must build with safety and security in mind from day one. Protecting your users, your customers, and your brand is paramount. Our frontier models are built on a foundation of rigorous AI principles and secure-by-design infrastructure.
Sustainability: In an era of rising compute demands, we must scale in a way that is both financially viable and positive for our planet. At Google, we are committed to carbon-free energy 24/7, ensuring that the UK’s AI growth does not come at the cost of our climate goals.

Architecting the future together

Google Cloud is the primary partner for the UK’s agentic transition. We are moving beyond the hype of experimentation into the rigor of production. From the research labs of King's Cross to the diverse enterprises powering the high street, we are architecting a resilient, sovereign, and prosperous future for the United Kingdom.

Thank you to everyone who’s joining us in London — yesterday, today, and into the future. This year we’ve packaged up an exclusive on-demand experience, allowing you to stream the defining London Summit moments, available anywhere, anytime.

How Atlas scales hundreds of merchant databases with Cloud SQL Enterprise Plus edition

Tue, 16 Jun 2026 16:00:00 +0000

Atlas is building the operating system for restaurants. Online storefronts, point of sale, third-party logistics, food platform integrations, customer loyalty, and AI tools represent everything a restaurant needs to start, run, and grow. We work with brands like SaladStop, Killiney, Haidilao, Raffles Hotel, Lo and Behold Group and the Les Amis Group in Singapore, helping merchants increase basket sizes, grow sales, and reduce operational costs.

Every merchant on Atlas gets their own dedicated Cloud SQL for PostgreSQL database. Restaurants are very different from each other. A single-outlet cafe and a multi-outlet chain should not look the same underneath. Isolated databases give us full data separation, predictable performance even during peak lunch and dinner rushes, and the flexibility to scale, tune, or migrate each merchant independently. As Atlas grows, the number of databases grows with us.

The challenge: Scaling beyond standard

We started on the standard Cloud SQL Enterprise edition. It was a solid foundation, but as we onboarded more merchants and shipped more features, the operational layer required to manage our databases became a bottleneck.

We were managing connection pooling as a separate layer, which meant more services to run, secure, and monitor. When a query caused a CPU spike, we needed to know exactly what happened and which merchant triggered it, but we were spending too much time reconstructing problems from limited signals. With a lean team and no dedicated database engineers, every extra component multiplied the maintenance load.

The shift to Enterprise Plus edition

When we needed to provision new database instances, the Google Cloud team introduced us to Cloud SQL Enterprise Plus edition. We were already asking ourselves how much more operational overhead this was going to add, and what stood out was that Enterprise Plus edition removed whole categories of work we would otherwise have to own.

Managed connection pooling: Now built directly into Cloud SQL, we no longer run pooling as a separate layer. This means fewer moving parts, less to maintain, and a smaller security surface area.
Query insights: This was the most impactful feature for our needs. We can now see exactly which queries are expensive and which merchant is triggering them. It turns performance tuning from guesswork into something concrete and actionable. For a platform running hundreds of databases, this visibility is a "superpower."
Data cache: This keeps read performance consistent even as merchant datasets grow. Since restaurants generate more data every day, the data layer needs to stay fast as that complexity compounds.
Near-zero downtime scaling: We can now scale instances as merchants grow without disrupting service during off-peak hours.

After seeing the results on the new instance, we migrated all our existing databases to Enterprise Plus edition as well.

The impact: Focus on innovation, not plumbing

Atlas today powers thousands of restaurant outlets, processes tens of thousands orders daily using hundreds of managed databases. The biggest change is where engineering time goes. We spend 30% less time on database operations and more time building products. Merchant onboarding got simpler because a new merchant is provisioned in seconds with a ready-to-use managed database. We are much more proactive on performance now, catching and fixing issues before they reach merchants. Day to day, we are not thinking about database plumbing. We are thinking about how to serve merchants better and that has allowed Atlas to grow 200% to 300% year over year.

Looking ahead: An AI-first future

We are investing deeply in AI, both internally and externally. Internally, we have gone all in on agentic engineering through AI-assisted development workflows that let a lean team build, review, and ship code significantly faster. Externally, we are building AI-powered tools that help restaurant operators make better decisions and act on them. We have a lot of experimental ideas on the roadmap, including new product surfaces and new ways to help restaurants grow. The thing that gives us confidence to move fast on all of this is that the foundational layer, Cloud SQL and Google Kubernetes Engine (GKE), is battle-tested and does not get in the way.

Google Cloud handles the infrastructure complexity. Atlas stays focused on building the best tools for restaurants.

Cloud SQL Enterprise Plus gave us a database architecture that is flexible, observable, and easy to scale. We are not thinking about infrastructure anymore, we are thinking about our merchants. As we go deeper on AI and continue growing the platform, Google Cloud gives us the confidence to move fast without worrying about what is underneath.

Ready to scale your database architecture?

Don't let infrastructure bottlenecks slow down your innovation. Whether you are managing tens or hundreds of databases, see how Google Cloud SQL can streamline your operations, enhance observability, and give your engineering team the freedom to focus on what matters most.

Explore Cloud SQL Enterprise Plus edition today
Sign up to try Cloud SQL for free.

How Siemens "slices the elephant," advancing agentic workflows for industrial software development

Tue, 16 Jun 2026 07:00:00 +0000

For technology companies like Siemens, software is the nervous system of factories, energy grids, and transportation networks worldwide.

As a global leader in industrial AI, industrial software, and industrial automation, Siemens brings decades of domain expertise across factory and process automation, energy infrastructure, and intelligent transportation — expertise that no off-the-shelf AI solution can replicate. But innovation carries a heavy anchor: legacy code.

With codebases spanning hundreds of millions of lines developed for over more than a decade, Siemens faced a challenge that standard AI tools couldn't solve: understanding and modernizing this code and the applications which run on it. The scale and depth of industrial-grade software demand a fundamentally different approach. Existing coding assistants lacked the contextual depth required to navigate complex, multi-layered industrial codebases — a gap Siemens set out to close.

To solve this, Siemens and Google Cloud created Knowledge Fabric, an AI system for automating the software development lifecycle. It was built using knowledge graphs on Spanner Graph, the Google Agent Development Kit, Gemini API, Gemini Enterprise Agent Platform, Gemini CLI, and Anthropic Claude Code. In a pilot migrating existing frontiers to web-based interfaces, Knowledge Fabric reduced implementation effort, freeing engineers to focus on customer innovations while maintaining full system compatibility.

“By ingesting the entire software ecosystem into an intelligent agentic system equipped with custom knowledge graphs, we aren’t just helping developers optimize their development time; we are enabling autonomous agents to reason across the past to build the future,” said Franz Menzl, senior vice president, product creation excellence at Siemens. “This is about freeing engineers from repetitive work so they can focus on higher-value problem solving.”

The challenge: the complexity of industrial software

Modernizing large-scale industrial-grade software systems is often compared to rebuilding a jet while flying it. For Siemens, the challenge had four dimensions:

Scale: The repositories are massive — far exceeding the context windows of standard large language models.
Fragmentation: Critical knowledge was scattered across code, Jira tickets, Confluence pages, and scanned PDF manuals from the early 2000s.
Complexity: Tracing the link between a specific line of code and a functional requirement document from 10 years ago presented a challenge that no manual or conventional tooling approach could address efficiently. It’s a reality shared across the industry.
Responsibility: Systems must adhere to strict quality, compliance, and lifecycle requirements, often over 15 to 20 years of operation. AI‑generated outputs must therefore be explainable, traceable, and verifiable. Hallucinated or unvalidated changes are not merely inefficient but operationally unacceptable.

"We realized that standard RAG (retrieval-augmented generation) wasn't enough," said Agata Gołębiowska, technical lead, Google Cloud. "Code isn't just text; it has inherent structure. A class belongs to a file, which belongs to a module. Flattening that into a vector database meant losing the representation of relationships elements of the codebase."

The solution: A domain-aware Knowledge Fabric

To make this sprawling software environment navigable for AI-driven workflows, the teams built the Knowledge Fabric agent. This agent goes beyond keyword matching to “understand” the relationships between assets.

We use Spanner Graph to model the inherent structure of the codebase, applying the same rigor to documentation across formats. By mapping connections between these domains, we can link specific code snippets directly to requirements in a design document. Agents then traverse this graph, using tools to query the structure via Graph Query Language (GQL).

But GQL is only one piece. To enable semantic understanding, we generate embeddings for every node, using Spanner's Approximate Nearest Neighbors (ANN) algorithm to perform efficient vector search across the full codebase. Finally, we give agents full-text search capabilities, which can be combined with GQL to pinpoint nodes and edges with precision.

Combining these three methods lets an LLM agent answer complex queries, such as: "Which functions need to be updated if I change the logic in the Axis Control Panel?" The system traverses the graph — weighing keyword and semantic similarity — to identify dependencies, retrieve relevant documentation, and present a precise impact analysis.

This precise context is what lets a coding agent produce a valid, usable, and maintainable implementation.

"Slicing the elephant:" the agentic workflow

A key insight from the project was that AI agents struggle with massive, ambiguous tasks. To succeed, the team adopted a design pattern dubbed "slicing the elephant."

The system breaks a sweeping request like “refactor this module” into smaller, more manageable tasks, each handled by a specialized agent built with the Google Agent Development Kit (ADK):

Search agent: Acts as a deep-research specialist. It uses tools to explore the code graph and cross-reference findings with documentation in Agent Search.
User story agent: Interviews the product owner to gather requirements, then drafts detailed user stories with acceptance criteria linked to existing system contexts.
Architecture impact agent: Analyzes proposed changes against the graph to predict side effects before a single line of code is written.
Task breakdown agent: Consumes the analysis from the architecture impact agent and breaks the work into small, manageable tasks, each carrying all the context relevant to a specific change.
Coding agent: Implements the change described in a specific task. Reaching this step without context and prior analysis produces unusable code.

The system keeps a human in the loop at every step, which ensures reliable, production‑grade outcomes and keeps engineers focused on meaningful work rather than routine implementation.

"By slicing the elephant — breaking complex refactoring jobs into smaller, agent-led tasks — we observed a significant productivity increase," said Alexander Lomakin, project lead at Siemens. "We essentially gave the AI the roadmap it needed to navigate the complexity."

Pilot results: Faster, more efficient engineering

Developers saw results almost immediately.

Analyzing dependencies for a new feature once required senior engineers to spend several days navigating codebases and legacy documentation. With the Knowledge Fabric, the same work now takes far less time.

In a recent production pilot migrating legacy control panels to modern web‑based interfaces, the Knowledge Fabric reduced overall coding effort while preserving system integrity and industrial quality standards.

Engineers now spend more time creating customer value and less on repetitive work.

Get started

The Knowledge Fabric shows that generative AI can do more than write boilerplate code, it can also help teams modernize the legacy systems their businesses depend on most.

To learn more about building graph-based agents for your own legacy modernization:

Read about Spanner Graph.
Explore Agent Platform and find pre-built production-grade agents on Agent Garden
Check out the Agent Development Kit.
Read more on how Siemens is advancing industrial AI.

Architecting a trusted agentic platform with graph technologies: A Yahoo case study

Mon, 15 Jun 2026 16:00:00 +0000

As enterprises adopt agentic AI, they need to shift from reactive systems of intelligence to proactive systems of action to equip the agents they’re building with the context and performance they need, plus regulator-grade accountability, where every decision is explainable and auditable.

At Google Cloud Next ‘26, we discuss how our Agentic Data Cloud enables a system of action, and Yahoo’s digital media buying platform is a compelling example of this vision. Yahoo partnered with Google Cloud to build its Seller Agent digital media buying platform using Google Data Cloud graph technologies. Seller Agent condenses multi-week manual processes into fully governed, live campaigns that can be executed in just seconds. Ultimately, this agentic platform serves as a powerful blueprint for multiple industries, demonstrating that autonomous systems can operate at remarkable speed while remaining strictly accountable.

"Yahoo's mission is to be a trusted guide through the digital world. In partnership with Google Cloud, we're extending that promise to advertisers: agentic media buying that's fast, transparent, effective, and built to be trusted." - Gabriel DeWitt, Head of Monetization, Yahoo

In this blog, we explore the shift toward agentic AI, examine how Yahoo’s Seller Agent architecture solves for speed and trust in media buying, and show you how to apply this graph-based pattern to build trusted systems of action in your own organization.

Case study: agentic media buying

For years, complex, high-value workflows—like premium digital advertising campaigns—have required weeks of human handoffs, fragmented spreadsheets, and manual analysis. Yahoo recognized that agentic AI could collapse this timeline, allowing agents to plan and execute campaigns in mere seconds. This leap from manual to autonomous execution represents a massive opportunity to reclaim operational efficiency and ensure more of every dollar reaches measurable outcomes.

But simply dropping LLMs into a high-stakes workflow does not solve the problem; an agent attempting to negotiate contracts or ad placements without a deterministic understanding of real-time inventory, pricing rules, and business constraints is prone to hallucinate — potentially resulting in disastrous deals. A trusted agentic platform requires a definitive, real-time source of truth, ensuring it acts on hard facts rather than statistical guesses.

Furthermore, speed and factual grounding are only half the equation. The moment an AI agent starts moving real budgets, it faces scrutiny from regulators who demand instant answers to why specific decisions were made or which policies were applied. Digging through raw system logs after the fact is the wrong control surface for autonomous execution. Real-world systems of action require regulator-grade governance and auditability built directly into the workflow, not bolted on as an afterthought.

The architecture of a trusted system of action

Yahoo's mission has always been to be a trusted guide through the digital world. Agentic media buying extends that promise to advertisers, agencies, publishers, and regulators who entrust Yahoo with their budgets — and expect real accountability. The issue was automating campaign execution in a way that was explainable, governable, and auditable.

To meet this challenge, Yahoo built its Seller Agent as a multi-agent system running on Google Cloud. Buyer requests enter through a planning supervisor agent running on Google Kubernetes Engine (GKE) and orchestrated with Google's Agent Development Kit (ADK). The supervisor decomposes each request into specialized tasks including inventory discovery, audience matching, forecasting, pricing analysis, package recommendation, governance review, and execution. Agents coordinate through the open Agent2Agent (A2A) protocol, while Gemini Enterprise Agent Platform hosts models for embeddings, forecasting, and graph learnings.

But the true breakthrough — what makes autonomous execution both fast and fully transparent — is the platform’s dual-graph foundation. The platform is anchored by two specialized graph systems with an intentional separation of duties: a knowledge graph that’s optimized for acting, and a second context graph for remembering and learning.

"As the industry moves from systems of intelligence to systems of action, the constraint on autonomous AI shifts from model capability to whether a business can trust what an agent does unsupervised. Autonomous systems must record why decisions were made and learn from outcomes. That trust is earned through robust data infrastructure. We built that foundation with Google Data Cloud: a knowledge graph for operational truth in Spanner Graph, a context graph for decision lineage in BigQuery Graph — the blueprint for enterprise-scale agentic platforms." - Swapnil Patel, Senior Director and Head of Monetization Engineering, Yahoo

The knowledge graph: Grounding agents in business reality

Powered by Spanner Graph, Yahoo’s knowledge graph represents its monetization business as a connected operational model, grounding every agent decision in business reality. It models advertising products, placements, audience segments, inventory, contracts, and governance controls as first-class entities and relationships. Crucially, policies live directly within the graph as versioned relationships rather than being buried in application logic. This design allows the system to evaluate products, contractual obligations, consent requirements, and regulatory constraints together in a single, unified graph traversal.

The graph acts as a semantic contract across the agentic platform. During campaign evaluation, an agent can navigate from initial buyer requirements to eligible audiences and governing policies within a single query plan. Gemini Enterprise Agent Platform embeddings enrich these entities with semantic similarity, while graph neural networks contribute inferred relationships. Ultimately, this allows agents to do more than just retrieve available inventory — they understand exactly why it is relevant and help ensure it satisfies all governing constraints.

Yahoo’s knowledge graph ontology, aligned with industry standards like IAB AdCOM

The context graph: creating an auditable memory

Execution at agent-scale is only safe if it is entirely transparent — which is the core function of the context graph. Every time the Seller Agent takes an action, that exact operational span is captured by the BigQuery Agent Analytics plugin. In addition to logging the raw events, the system shapes this evidence into a typed, queryable context graph using BigQuery Agent Analytics SDK utilizing Yahoo's decision-trace ontology, stored in BigQuery Graph.

Consequently, every decision point, candidate package, policy evaluation, specialist-agent delegation, and execution outcome becomes a connected graph of evidence. Because this trace is structured as a typed graph, explaining the agent’s decision making process becomes a simple query. An auditor can instantly trace a decision from the originating campaign brief through every score that’s assigned and policy that’s applied. This transforms autonomous behavior from an opaque process into a fully transparent and continuously improving record of decision-making, helping to ensure absolute accountability.

Yahoo’s context graph ontology

From human to agent scale

For a concrete example of the architecture in action, consider an ad campaign run. What traditionally required weeks of coordination across planning, sales, operations, and compliance can now be completed in seconds through two simultaneous processes.

Acting via the knowledge graph. This pipeline moves the budget, navigating linearly from the buyer's request to a live campaign ground on the knowledge graph. This proceeds in four steps:

Submitting the brief: A buyer agent submits a campaign brief over Ad Context Protocol (AdCP) that describes the desired audience, budget, geography, and business objective.
Knowledge retrieval: The Seller Agent queries the knowledge graph to identify relevant inventory, audiences, contractual availability, historical performance, and governing policies.
Evaluation and scoring: The agent evaluates these factors together to assemble a package of media buying candidates. Forecasting models score the opportunities, while a governance agent independently reviews consent, brand safety, and regulatory constraints.
Approval and execution: The package is either approved automatically under policy thresholds or escalated for human review. Once approved, the media buy is executed and activated.

Auditing and learning via the context graph. While the execution pipeline moves forward, this parallel loop continuously captures the system's reasoning in the context graph, helping to ensure transparency and improve future cycles. This offers the following capabilities:

Continuous capture: Every candidate considered, score assigned, policy applied, and governance decision becomes a connected record in the context graph, linked to the originating campaign session.
Closed-loop learning: As delivery, attribution, and outcome signals arrive, they are joined back to the decisions that produced them, creating the training data that improves future recommendations.
Instant explainability: If an advertiser asks why a particular package was selected or which policies influenced the outcome, the answer is preserved in the context graph and reachable through a single query.

The result is a platform where knowledge, decision-making, governance, measurement, and learning operate together — allowing autonomous media buying to remain explainable, auditable, and continuously improving.

A blueprint for many industries

The era of AI as a mere advisor is ending. Enterprises are demanding systems of action — autonomous agents capable of executing complex, multi-step workflows. But in regulated sectors, the speed that AI brings to the table turns into a liability if you cannot prove how a decision was made. The primary barrier to autonomous execution is no longer intelligence; it is trust.

The architecture that Yahoo and Google Cloud built provides a broadly applicable blueprint with which to solve this. While designed to fix the bottlenecks of digital media buying, the underlying pattern applies to any industry managing high-stakes decisions — from financial trading to supply chain logistics. To operate at agent speed but still maintain human oversight, enterprises must adopt a new architectural baseline that:

Grounds decisions in business reality: Agents cannot rely on probabilistic models alone. They must be grounded by a knowledge graph that deterministically maps your business logic, active contracts, and compliance rules.
Builds an auditable memory: You cannot govern what you cannot trace. Every agentic action must be captured in a context graph, creating an immutable, queryable record of exactly why a decision was made and which alternatives were rejected.
Embraces open interoperability: Trust requires transparency. By building on open protocols and provenance standards, industries can establish a common, auditable language for agentic behavior.

As foundational models become commoditized, enterprises’ competitive advantages are shifting. Long term, your moat will not be the language model you deploy, but the proprietary graph of your business operations and governed history. Likewise, the future of enterprise AI isn’t simply systems that can act, but systems that can explain, govern, and take accountability for those actions.

Get started today

Ready to build your own trusted system of action? Start by exploring Spanner Graph to ground your agentic workflows in business reality. Next, use BigQuery Graph to build an auditable memory that powers closed-loop learning and regulator-grade explainability. You can begin capturing and analyzing these operational traces today using the BigQuery Agent Analytics Plugin and SDK. Finally, review the Ad Context Protocol to understand the open communication standards underpinning Yahoo’s agentic platform.

Modernizing Healthcare: How Alcidion achieved greater stability and performance with AlloyDB

Mon, 08 Jun 2026 16:00:00 +0000

In clinical informatics, every second counts. For Alcidion, a global leader in smart health solutions, the mission is simple but critical: use technology to reduce cognitive load for clinicians and present the right information at the right time to save lives.

Whether it’s managing patient flow in an emergency department or ensuring a patient is in the correct ward to avoid adverse outcomes, Alcidion’s flagship platform, Miya Precision, serves as a dynamic intelligent care platform for modern hospitals. To power this mission, the platform recently underwent a major architectural transformation, migrating from a legacy Microsoft SQL Server environment to Google Cloud’s AlloyDB for PostgreSQL.

The challenge: overcoming performance bottlenecks

Operating in an industry where data integrity and uptime are non-negotiable, Alcidion faced several technical and operational hurdles with its previous setup:

Operational overhead: Managing persistent backends for SQL Server required significant manual effort. The team had to manually balance database loads between elastic pools to maintain performance while trying to optimize costs. They also had to constantly manage the gap between allocated and used space to prevent shared pools from being consumed by excessive slack space.
Performance latency: Complex JSON data processing, critical for modern health informatics, was taking up to 30 minutes for certain jobs.
Stability concerns: The team sought a more stable Kubernetes environment and a persistent backend that could scale without constant administrative intervention.

The solution: a smooth migration to AlloyDB

Alcidion used the Database Migration Service (DMS) to move from SQL Server to AlloyDB, achieving a remarkably efficient cutover. The total learning and migration process took under one month, with the core database move completed in only one and a half weeks.

By creating custom synchronization tools and using Google Cloud’s managed services, the team reduced the final transition window to just 15 minutes. Alcidion achieved this by spinning up a new Google Cloud instance synchronized to the active one, with both accessible via unique fully qualified domain names. The new environment remained in read-only mode for customer validation.

During the final cutover, the old instance was set to read-only, synchronization was halted, and external integration links were toggled to the new environment. This streamlined process allowed users to log into the new instance and resume work within minutes, with the primary delay being DNS record updates.

Alcidion chose a fully managed AlloyDB service to eliminate control plane tasks and administrative overhead. This shift allows their engineering team to focus on clinical innovation and product development rather than "managing the container" or the underlying database infrastructure.

Being able to cut over to AlloyDB in about 15 minutes had our users back to work almost immediately. For a system clinicians rely on around the clock, that kind of smooth transition gave Alcidion real confidence.

The results: impact by the numbers

The shift to AlloyDB and Google’s Agentic Data Cloud has delivered immediate, quantifiable improvements for Alcidion and its healthcare customers:

Faster data processing: Data processing that previously relied on SQL Server stored procedures — a process that became increasingly time-consuming as data volumes grew — has been transformed. By migrating to AlloyDB and using BigQuery and Dataflow for processing, Alcidion has seen jobs that once took 30 minutes now complete in just 5 to 60 seconds.
Enhanced stability: The migration has delivered a step-change in reliability. In the previous environment, the team faced monthly disruptions, ranging from failed scheduled maintenance to connectivity issues that required manual intervention. In contrast, AlloyDB and Google Cloud’s compute services have proven exceptionally stable, allowing the team to move away from the "firefighting" mode associated with frequent infrastructure crashes.
Reduced cognitive load: By simplifying their backend and clinical dashboards, Alcidion’s SREs have significantly reduced their administrative burden. This shift has freed the team to focus on high-value innovation, such as refining predictive analytics and generative AI that empower clinicians to make informed clinical decisions faster.

Future vision: AI and beyond

Alcidion isn't stopping at database modernization. The move to AlloyDB is a foundational step for their next phase of growth:

AlloyDB columnar engine: The team is exploring the columnar engine for a second round of query optimization and real-time analytics.
Generative AI apps: Alcidion is actively working with Google to use AlloyDB’s Gemini Enterprise Agent Platform integration to perform concept analysis and pick out critical clinical insights from vast datasets.

By moving to AlloyDB, Alcidion has improved its stability and performance and built a strong foundation to keep delivering smarter, safer care to hospitals worldwide.

Ready to modernize your database? Learn more about how AlloyDB can transform your operational workloads.

How Trustpilot built a real-time architecture for data enrichment using Gemma

Mon, 01 Jun 2026 16:00:00 +0000

Processing millions of user reviews in real-time, under strict latency and cost constraints, is no easy task. Trustpilot has been doing exactly that with custom machine learning since long before large language models (LLMs) were cool. Now, as the company transitions its core stack to generative AI, here is a look at how we teamed up to build a high-volume streaming pipeline using fine-tuned Gemma models.

Powering deep review intelligence at scale

Trustpilot’s core business relies on delivering deep, actionable review intelligence. As a platform championing transparency and genuine feedback, it must safeguard data integrity and maximize value. This means extracting every drop of metadata from incoming reviews — making LLMs the perfect tool for the job.

These models excel at parsing messy, human-written text to run named entity recognition (NER), categorize business domains, score sentiment, and pinpoint customer intent. But while prompting an LLM for a few reviews is easy, processing millions in real-time without blowing up costs is a massive engineering hurdle.

Why fine-tune an open model?

When pursuing such a big task, why isn’t just plugging into a powerful, off-the-shelf, frontier model like Gemini the right approach? For a pipeline this critical to the core business, closed models are rarely the best option. Instead, by fine-tuning open-weight models like Gemma, Trustpilot takes full ownership of their AI strategy. Here’s how:

Total model independence: By owning its models, Trustpilot ensures it controls the retraining lifecycle, completely freeing it from a third-party vendor's update schedule or sudden API changes.
Predictable economics: Shifting from a variable per-token pricing model to fixed infrastructure costs makes running millions of predictions financially viable and optimizable.
Expanding MLOps capabilities: Building these models in-house enables Trustpilot to bake in the "secret sauce" of its review intelligence while building competencies on open-weight models.
Architectural continuity: Standardizing on an open-weight lineage preserves the company’s ability to leverage the future iterations of the base model. This enables performance gains with minimal engineering overhead.

Rather than deploying one massive model, Trustpilot built a suite of highly specialized models using the lightweight google/gemma-2-9b as a base.

To get heavy-weight performance from a small footprint, the company employed a consensus annotation over a stratified sample of the Trustpilot review corpus, using a selection of teacher models from the Gemini 2.0/2.5 Pro/Flash family. This process generated high quality training datasets for specialized tasks like topic classification, NER, and sentiment extraction.

The datasets were subsequently used to fine-tune a targeted lineup of custom models that considerably outperformed the legacy solution and delivered accuracy just a couple percentage points lower than the teacher models’ consensus.

System architecture

This architecture was built on top of Dataflow and Gemini Enterprise Agent Platform Endpoints, which play together very nicely because of the out-of-the-box VertexAIModelHandlerJSON.

We decoupled business logic and raw LLM inference by creating two separate endpoints:

The classifier: a FastAPI-based endpoint that handles the messy stuff, pre/post-processing, prompt templating, and chaining.
The LLM: A separate Agent Platform endpoint dedicated strictly to serving the Gemma model via vLLM.

This approach keeps the Dataflow job clean and ensures the LLM endpoint sticks to what it does best: generating text. Plus, it allows Trustpoint to scale them independently based on the traffic.

Performance tuning

To get the most out of the vLLM-based Agent Platform endpoints, Trustpilot focused on squeezing every bit of performance out of the entire pipeline, especially from the A2 VMs using A100 GPUs. It also leveraged the customized and optimized version of vLLM maintained by Gemini Enterprise Agent Platform.

A focus of our performance tuning involved optimizing the vLLM backend configuration to prevent processing bottlenecks. By carefully adjusting the engine parameters, selecting the appropriate data type, and enabling useful settings such as prefix caching, we ensured the models could smoothly handle high streaming volumes.

Together, we also created a reusable load testing framework to find the optimal serving capacity for a vLLM inference server and to sketch its performance profile. This enabled setting a baseline for needed infrastructure, and tuning the auto-scaling setup using the request count-based metric. In addition, a new metric using vLLM number of requests waiting could be even better for this.

Challenges

While building this setup, Trustpilot encountered a few notable hurdles:

Private networking: The architecture aimed to be fully isolated by using private endpoints and Private Service Connect, but this wasn’t possible because there was no native support for direct private communication between distinct endpoints.
Deployment observability and reliability: Endpoint deployments can be slow or opaque, which occasionally requires extra troubleshooting when entering an unhealthy state. Trustpilot is still working closely with the Gemini Enterprise Agent Platform product team to help shape future observability features and platforms.
GPU Scarcity: Securing A100 GPUs in the EU region is tough, so on-demand VMs are often a no-go. Instead, leveraging reservations is preferable but balancing them between development, production, training, inference, and experiments can be quite challenging.

The results

Together with Google Cloud, Trustpilot leveraged the full potential of Gemma on Gemini Enterprise Agent Platform to process millions of reviews a day in near real-time. In doing so, they achieved Gemini-like performance for a fraction of the cost. This ultimately allowed the Trustpilot Business Platform to turn millions of everyday customer reviews into instant, actionable insights. You can read more on the Trustpilot Medium blog post.

^{This blog post was written by Assulan Nurkas (Trustpilot), Subu Ramasubramanian (Trustpilot), Konrad Stanek (Trustpilot), Dario Banfi (Google) and Michael Cohen Hjertén (Google) based on the work done during the joint project at the end of 2025.}

Cool stuff Google Cloud customers built, May edition: Agentic algorithms for supply chains; virtual try-on APIs; robotic camera operators & more

Fri, 29 May 2026 16:00:00 +0000

AI and cloud technology are reshaping every corner of every industry around the world. Without our customers, who are building the future on our platform, there would be no Google

Cloud. In this regular round-up, we dive into some of the exciting projects redefining businesses, shaping industries, and creating new categories.

For our latest edition, we learn how Urban Outfitters sped up its order management; BASF uses AlphaEvolve algorithms to map global supply chains; the unification strategy for UKG’s workforce intelligence; WPP’s secrets to training humanoid robot camera operators; how Breuninger piloted Virtual Try-On APIs; creating automated video clips with Glance; and Movix improves the production of dental aligners.

Be sure to check back next month to see how more industry leaders and exciting startups are putting Google Cloud technologies to use. And if you haven’t already, please peruse our list of 1,302 real-world gen AI use cases from our customers.

Urban Outfitters saves big by migrating order management

Who: Urban Outfitters, Inc. (URBN), the popular clothing and home goods retailer, relies on IBM Sterling OMS as the nerve center of its global ecommerce operations. However, the foundation of this critical system — a massive 11TB Oracle database — was increasingly becoming a bottleneck.

What they did: URBN completed a major infrastructure upgrade, migrating its IBM Sterling OMS from an Oracle database to Google Cloud's AlloyDB for PostgreSQL. To enhance performance and provide high availability and scalability, the AlloyDB deployment architecture includes two read replicas, providing low-latency access to data for reporting and analytics. Google Cloud and IBM teams also assisted URBN in a rigorous, iterative switchover testing strategy.

Why it matters: The migration to AlloyDB has fundamentally reshaped URBN’s data strategy, delivering a more favorable total cost of ownership through an optimized storage and compute architecture, without sacrificing performance or reliability. Furthermore, the shift to a PostgreSQL-compatible database gave URBN the flexibility of an open-source ecosystem, providing freedom from vendor lock-in, as well as significant speed improvements that enhanced responsiveness.

Learn from us: "URBN’s successful migration serves as a blueprint for organizations looking to modernize their mission-critical infrastructure and future-proof their environment for AI expansion. This journey proves that even the most complex, mission-critical migrations can be achieved through deep cross-organizational partnership and a phased, risk-mitigated approach." – Rob Frieman, CIO, Urban Outfitters & Raj Pai, VP, Product Management, Databases, Google Cloud

BASF manages supply chain decisions with AlphaEvolve

Who: BASF Agricultural Solutions manages a complex network of 180 production sites with more than 5,000 distinct value chains. Currently, human planners make thousands of local decisions every day on what to produce, when to produce it, and how much safety stock to hold.

What they did: To understand how local decisions ripple across their entire global network, BASF turned to AlphaEvolve on Google Cloud to build a digital twin of their supply chain. In collaboration with Google Cloud and prognostica GmbH, BASF fed the model three years of historical data and then generated variations of the code, mutating the logic to see if it could simulate a supply chain that matched the real-world historical data.

Why it matters: By running thousands of experiments, AlphaEvolve developed a clear, human-readable algorithm that explains how the BASF network truly operates. The final algorithm successfully mirrored the actual historical performance of the supply chain, significantly reducing the error rates compared to the initial seed model. It automatically discovered factually correct, domain-specific supply chain rules, providing a clear foundation for optimizing asset utilization globally.

Learn from us: “We had several attempts to build a digital twin. … By using AlphaEvolve, we cannot only map the complex network based on system data, but at the same time understand and copy the human decisions that drive our daily operations.” – Dr. Goetz Krabbe, vice president for global supply chain at BASF

UKG unlocks real-time workforce intelligence at scale

Who: UKG is one of the leading providers of human capital management (HCM) and workforce management (WFM) solutions, but years of growth led to backend sprawl. They have 126 application teams, dozens of tech stacks, and more than 12,000 database instances.

What they did: To bring the full UKG suite onto one real-time foundation, the company built People Fabric, a new data and intelligence platform powered by AlloyDB for PostgreSQL and the just-announced Agentic Data Cloud. They created a custom change data capture (CDC) framework to extract changes from existing operational databases, and for larger analytical workloads, the same data flows into BigQuery, while Cloud SQL holds the metadata and tenancy context.

Why it matters: People Fabric gives UKG a complete and consistent view of people, work, pay, and culture data that’s updated continuously and ready for AI to use in real time. For engineering teams, People Fabric acts as a database-as-a-service that accelerates development and supports modernization without customer disruption. Additionally, migrating core person and employment data off their on-prem monolith has generated cost savings significant enough to fund half of People Fabric.

Learn from us: “As we continue expanding People Fabric, we’re laying the groundwork for deeper agentic automation, more responsive analytics, and a growing set of AI-driven capabilities — all on a trusted, scalable foundation built for what’s next.” – Radhi Chagarlamudi, Group Vice President, Product Engineering, UKG & Heather White, Cloud Data Architect, Google Cloud

WPP accelerates humanoid robot training 10x with G4 VMs

Who: WPP is one of the world’s largest marketing organizations, handling $70 billion of media for enterprise clients. They work on some of the most complex commercial film shoots and were eager to test the viability of robotic cameras to capture more footage, but this required complex training of physical models AI.

What they did: WPP used the new G4 VM instance powered by NVIDIA RTX PRO 6000 Blackwell on Google Cloud to tackle the unique challenges of training physical AI for robotics in videography settings. After capturing human motion with the OptiTrack mocap system, they undertook reinforcement learning using the AI Hypercomputer together with the NVIDIA Isaac Sim image. MuJoCo, an open source physics engine by Google DeepMind, was a critical piece of simulation software that validated accuracy continuously, in real-time.

Why it matters: WPP was able to utilize a P2P topology that moves data directly between GPUs without the bottleneck of central processing. They saw speed increases in excess of 10x, taking training times down to less than one hour. Through high-volume simulation, the humanoid robots learned how to respond to small changes and bridge the tough "sim-to-real" gap, helping ensure the robot's simulated adaptability translated to safety and stability in the real world.

Learn from us: "Our process for mastering complex, natural movement on a film set can be replicated across industries to overcome the massive computational complexity of training robots." – Perry Nightingale, SVP of Creative AI, WPP

Breuninger boosted sales with its "be your own model" AI

Who: Breuninger, a fashion and lifestyle company based in Germany, thought emerging generative media models could be a good fit to answer the question every online fashion shopper asks: "How will this look on me?"

What they did: Working with Google Cloud, they built a virtual try-on experience that lets shoppers see high-end fashion on their own bodies using a simple selfie. Using the Virtual Try-On (VTO) API, Breuninger’s data team worked directly with Google’s engineers to test and refine the technology in three stages, ultimately moving from pre-selected models to a user-first, selfie-based approach. The project was also part of Breuninger’s move to a Flutter-based platform, which helped the team move from its vision to a live launch in only three months.

Why it matters: During a six-week A/B test over Black Week and the holiday season, the team found that shoppers who used the virtual try-on converted purchases at a higher rate than those who didn't. Customer surveys reinforced the numbers: shoppers responded well to the high image quality and the personalized experience.

Learn from us: “Breuninger continues to refine the experience based on how customers actually use virtual try-on in everyday shopping — the same user-first approach that shaped the project from the start.” – Daniel Rascher, Senior Product Owner, Breuninger & Dr. Michael Menzel, Customer AI Specialist, Google Cloud

Glance turns hours of video into mobile-ready clips

Who: Glance, a mobile-first content platform, processes 1-2 hour videos from sources like podcasts, news reports, movies, and web series, and transforms them into 30 to 180-second vertical clips optimized for mobile lock screens.

What they did: The goal was to create a complete pipeline that takes a long-form landscape video (16:9) and outputs multiple ready-to-publish short-form portrait videos (9:16). The final technical solution uses Google Cloud Speech-to-Text v2, Gemini, and the Google Vision API, combined with custom video manipulation using Samurai (an open-source object tracking tool), OpenCV and MoviePy. The process involves audio extraction, speech-to-text transcription, and using Gemini 2.5 Flash to analyze transcript text and identify optimal start and end timestamps for short video clips.

Why it matters: With daily volume projected to grow from 3,500 to over 10,000 videos per day, manual editing wasn’t a realistic path forward. Glance’s video pipeline demonstrates what becomes possible when AI handles the repetitive, judgement-intensive work of video editing. The system transforms thousands of long-form videos into mobile-ready clips each day, preserving narrative context while optimizing for vertical viewing. Rather than choosing between scale and quality, automated pipelines can deliver both.Learn from us: “Glance’s video pipeline demonstrates what becomes possible when AI handles the repetitive, judgement-intensive work of video editing. … The approach offers a template for any organization sitting on long-form video archives. Rather than choosing between scale and quality, automated pipelines can deliver both.” – Himanshu Aggarwal,

Machine Learning Engineer, Glance & Sharmila Devi, AI Consulting Lead, Google Cloud

Movix fills a gap in dental skills with specialized agentic AI

Who: Movix is building one of the first agentic AI solutions for dental appliance manufacturers and dental labs, to help solve a serious shortage of skilled dental technicians in aligner manufacturing.

What they did: Movix developed custom models for deep learning, computer vision, and 3D mesh analysis over a five-month period, using Google Cloud infrastructure. Once defects are detected, they use the Gemini Enterprise Agent Platform to generate client-facing feedback that reads as if it came directly from a human technician. Their 3D models use Cloud Run with L4 GPUs for the massive compute power required, and they use Compute Engine VMs to run experiments and train models.

Why it matters: Movix’s agentic solutions automate data entry and quality control, which are traditionally manual, time-consuming, and error-prone tasks. The automation and higher level of accuracy the QC agent delivers can save $300 per remake for an aligner manufacturer, and speed up the appliance manufacturing process with quicker turnaround times.

Learn from us: “We plan to build hybrid solutions … designing an architecture that connects our cloud-based AI agents with older, on-premises software that many conservative labs still use — through lightweight local connectors and standardized APIs. This will allow us to access a large market segment that has not yet migrated to the cloud.” – Marina Domracheva, CEO, Movix & Bakit Dzhumagulov, CTO, Movix