Garage Copilot — Retrieval-Augmented Repair Assistant
A question-answering assistant for DIY vehicle repair, grounded in a licensed document corpus with citations back to source.
- Implemented semantic retrieval in PostgreSQL using pgvector, returning ranked passages joined to their source metadata so every answer carries a verifiable citation — a question phrased as a symptom retrieves the right procedure with no keyword overlap.
- Added a tool-calling loop so the model queries the NHTSA recall API mid-answer, with the prompt and error path distinguishing a failed lookup from a genuine no recalls found — so it reports uncertainty rather than implying a vehicle is safe.
- Streamed LLM responses over server-sent events with client-disconnect handling, so abandoned requests stop generating instead of billing for tokens nobody reads.
- Deployed on AWS EC2 behind Caddy with automatic TLS; secrets loaded from AWS Secrets Manager into the container environment at start and never written to disk; no inbound SSH, with access via SSM Session Manager.
- Designed the corpus around licensing constraints — public-domain government data and manufacturer documentation under attribution — rather than scraping forum content, with attribution enforced at the schema level.
