Second Brain: Conversational RAG Chatbot

Built a local-first RAG chatbot that ingests 4,954 personal conversations into a 312K-chunk pgvector knowledge base. Improved retrieval recall from 0.58 to 0.85 on a 50-question eval set through context engineering. Runs LLM ops on AWS Bedrock with model-tier routing to balance cost and quality. Evaluation pipeline uses LLM-as-Judge for automated regression testing across retrieval and generation quality.