How the Amazon.com Catalog Team built self-learning generative AI at scale with Amazon Bedrock

This article describes how Amazon's Catalog Team built a self-learning generative AI system using Amazon Bedrock to automatically improve product attribute extraction and content generation at scale.

• Multiple smaller models process products through consensus; disagreements trigger supervisor investigation
• Supervisor agent resolves disputes and generates reusable learnings stored in dynamic knowledge base
• Learnings injected into worker prompts reduce future disagreements without retraining
• Post-inference signals from sellers and customers feed into same learning pipeline
• System uses Amazon Bedrock, Bedrock AgentCore, EC2, DynamoDB, SQS, and CloudWatch
• Disagreement rates serve as primary health metric; declining rates indicate effective learning
• Two deployment strategies: learn-then-deploy for new use cases, deploy-and-learn for established ones
• Architecture works best for high-volume, quality-critical, evolving domains

The system demonstrates how AI can accumulate domain-specific knowledge through production usage, continuously improving accuracy while reducing costs through intelligent model selection and learning capture.