Zalando builds an LLM proxy, CLI and risk-based PR approval bot for agentic engineering across 250+ engineering teams
ZalandoAmazon BedrockZalando's ML platform team deployed a LiteLLM-based API proxy in January 2024 giving engineers access to OpenAI, AWS Bedrock and Google Vertex models, plus a chat UI, a pydantic-ai CLI, a shared agent skill collection and governance via its Tech Radar. A risk-based PR approval bot auto-approves the 33% of PRs rated low-risk, which the source says reduced PR lead time by 20-40% compared with all PRs.
Overview
Zalando's ML platform team deployed a LiteLLM-based API proxy in January 2024 giving engineers access to OpenAI, AWS Bedrock and Google Vertex models, plus a chat UI, a pydantic-ai CLI, a shared agent skill collection and governance via its Tech Radar. A risk-based PR approval bot auto-approves the 33% of PRs rated low-risk, which the source says reduced PR lead time by 20-40% compared with all PRs.
This entry has 13 published fields tied to exact passages in an immutable source capture.
Inspect the highlighted sourceThe challenge
With more than 250 engineering teams exploring LLMs at different paces, Zalando needed API-based access to models from several providers, a way to measure adoption, governance and transparency across teams, and a way to protect PR lead time as AI coding increased PR sizes.
The solution
Zalando's ML platform team deployed a LiteLLM-based API proxy in January 2024 with access to models from OpenAI, AWS Bedrock and Google Vertex, complemented by a chat UI, a CLI built with pydantic-ai, and a centralized agent skill collection. A risk-based PR approval tool rates each PR as low, medium or high rollout risk at creation and auto-approves low-risk ones. Governance runs through an AI section of the Tech Radar.
Reported business value
The source states that 33% of PRs are low-risk and auto-approved by the bot, which in Zalando's case reduced PR lead time by 20-40% compared with all PRs. The proxy runs for 2k MAU on six small pods. Figures are self-reported.
Sources
Open any source and check the claim yourself — that is the point of the register.
This record was researched and written with AI assistance, and its claims were checked against the sources above. (EU AI Act art. 50 transparency notice.)
Other e-commerce entries in the register.
Foxintelligence Enhances Ecommerce Insight Using AWS and Generative AI
Foxintelligence by NielsenIQ, the largest ecommerce measurement and consumer analytics company in Europe with around 5 million active online buyers tracked, uses Amazon SageMaker to build, train, and deploy its machine learning models and is exploring Amazon Bedrock to access foundation models, fine-tune them, and do prompt engineering, reducing time to market for generative AI innovation. As a company handling consumer personal data, Foxintelligence relies on AWS security features to comply with the EU's General Data Protection Regulation (GDPR) and other regional regulations as it expands globally, with advice from the AWS engineering team on new compliance features.
ElasticRun helps small businesses improve supply chain with Databricks
Indian B2B e-commerce platform ElasticRun rebuilt its data infrastructure on Databricks with Delta Lake, Spark Declarative Pipelines and MLflow to manage over 10,000 machine learning models for supply chain and demand forecasting, cutting data pipeline slowdowns by 90% and IT costs by 33%.
Rakuten accelerates development with Claude Code
Rakuten, a large Japanese ecommerce company, uses Claude Code to automate coding tasks across its engineering teams, reducing time to market for new features from 24 days to 5 days and achieving 7 hours of sustained autonomous coding on a complex open-source refactoring project. Rakuten also deployed Claude Managed Agents across product, sales, marketing and finance, plugging into Slack and Teams for non-engineers to complete tasks.
Rakuten· JapanClaude Code · Claude Managed AgentsZalando Builds a GPT-4o Content Creation Copilot to Auto-Suggest Product Attributes
Zalando built a Content Creation Copilot that uses OpenAI's GPT-4 Turbo, later migrated to GPT-4o, to auto-suggest product attributes from images during article onboarding. The content creation stage was largely a manual process — copywriters used a Content Creation Tool and performed their own QA — contributing about 25% of the overall content production timeline. The system combines a prompt generator, article masterdata, and a translation layer mapping GPT output to Zalando's internal attribute codes, achieving roughly 75% accuracy while enriching about 50,000 attributes per week.
ZalandoOpenAI GPT-4 Turbo · OpenAI GPT-4oWas this helpful?
Your feedback helps us improve our use case database