← Projects

AI / RAG 2025 · Lead · team of 3

Relay

Retrieval-augmented assistant over internal docs and runbooks — answers with citations, not vibes.

EDIT: e.g. 40 pilot users EDIT: citation accuracy % EDIT: time-to-answer

Python LangChain Postgres OpenAI FastAPI

Problem

People were hunting through scattered docs, Slack threads, and tribal knowledge for answers that already existed somewhere — usually in a PDF nobody had opened in months.

Role

Led design and delivery end-to-end: retrieval quality, API, thin UI, and an evaluation set of real team questions.

Approach

Chunk and index the corpus, retrieve top matches, generate only from that context with explicit citations. Prefer “I don’t know” over a confident guess.

Architecture

Ingestion → embeddings + Postgres · FastAPI for query/retrieve/generate · simple ask-and-cite UI. Measured relevance and citation accuracy against a held-out question set.

Outcome

Shipped to a pilot group, tightened retrieval before expanding the corpus. Most of the hard work was data hygiene and eval — not the model call.

Lessons

RAG quality is mostly retrieval and content quality. Model choice matters less than clean chunks, good metadata, and an honest “no answer” path.