Built a RAG chatbot answering Peruvian tax questions from the SUNAT corpus. I was the only engineer on it: architecture, AWS Bedrock (Nova Lite), pgvector retrieval on PostgreSQL, the FastAPI backend, prompts, and admin panel. Load tested at around 1,000 queries a day. Raised answer accuracy from 60% to 92% on a labeled set of domain expert tax questions. I scored it with a mix of manual review and a second LLM as judge, then tuned the retrieval threshold and made the bot cite its sources. Fixed a production bug where the model answered from a user’s uploaded file instead of the official knowledge base. Added metadata tags so admin documents and per session uploads stay in separate scopes. Set up the AWS side: S3 for documents, Lambda for scheduled embedding refresh and notifications, CloudFront, and CloudWatch. Added websocket streaming so the first tokens arrive right away instead of after the full answer is generated.

Built the admin dashboard from scratch with React, Vite, and Material UI. It handles pet sitter approvals and is used by around 5 admins. Shipped two weeks before the deadline. I was the only junior engineer on the team.