Generative AI has shifted from experiments to production workloads like internal assistants and document summarization. Amazon Bedrock lets teams run models without managing GPU fleets. A minimal stack using API Gateway, Lambda, and Bedrock provides a secure, manageable starting point for early-stage projects, trading some flexibility for integrated IAM and AWS services.

The core flow is straightforward: an HTTP request hits API Gateway, triggers a Lambda function, which invokes a Bedrock model like Anthropic Claude Sonnet. For cost and security, start with broad IAM permissions in development, then restrict to specific model ARNs and regions in production. Logging requests to S3 or DynamoDB aids debugging and compliance.

Operational discipline is key. Manage prompts externally, enable Bedrock guardrails, and monitor costs via CloudWatch. For proprietary data, prefer retrieval-augmented generation (RAG) over fine-tuning by indexing documents in OpenSearch or a vector store. This approach keeps deployment simple while allowing for evolution as usage patterns stabilize.