Amazon FinTech builds AI system on AWS to automate regulatory inquiries

Amazon's FinTech team faced escalating challenges in processing regulatory inquiries across jurisdictions—knowledge fragmentation across thousands of documents in PDF, PPT, Word, and CSV formats; the need for multi-turn conversational context; and comprehensive observability to detect hallucinations or outdated guidelines. To address these, they built an intelligent regulatory response automation system on AWS. The solution uses Amazon Bedrock Knowledge Bases for RAG, with OpenSearch Serverless for vector storage, enabling precise retrieval from historical precedents. Real-time chat leverages Claude Sonnet 4.5 via the Converse Stream API, while DynamoDB manages conversation history for context-aware interactions.

Observability is a critical component, implemented through OpenTelemetry and self-hosted Langfuse to monitor retrieval accuracy, model decisions, and user interactions. The system avoids caching LLM responses due to the highly contextual nature of inquiries. An automated ingestion pipeline processes uploaded documents by generating pre-signed URLs through API Gateway and embedding data into the knowledge base. This approach allows each team to maintain its own dedicated knowledge base, scaling efficiently as inquiry volume grows. The result is a scalable, compliant AI application that reduces manual compilation time while maintaining rigorous standards for regulatory responses.