An internal assistant giving engineers source-grounded answers about a large multi-service platform.
Designed and built end to end — an internal assistant giving engineers source-grounded answers about a large multi-service platform. Agentic retrieval with tool use, a knowledge pipeline that keeps its knowledge current, and a category-balanced evaluation harness scoring correctness, completeness and faithfulness separately. Made answer quality a measured number rather than a feeling — and the honest number held back a wider rollout.
Python · FastAPI · Next.js · PostgreSQL · pgvector · Cloud Run · Claude / GPT / Gemini
The knowledge assistant's parts: a Next.js web app and a Python/FastAPI agent loop on Cloud Run. Each turn, the model chooses whether to search the PostgreSQL/pgvector knowledge store or grep and read source code. A knowledge pipeline keeps the store current, and a category-balanced evaluation harness scores correctness, completeness and faithfulness separately.Handed a fixed context, the model answers from whatever a single retrieval step returned. In the agent loop, a first knowledge-base search runs automatically; then the model decides each turn whether to search again, grep source or read a file, sees the results, and stops when it can answer or the turn limit forces an answer.