An AI research assistant that answers questions from a team's own documents, with citations. Scoped, built and launched in 19 days on Next.js and the Claude API — an MVP real users could judge before the next funding conversation.
The team had six months of runway and a thesis: their customers would pay for answers pulled straight from their own internal documents instead of a search box that returns twenty PDFs.
They needed something real enough to put in front of paying users — not a prototype, not a demo video — and they needed it before the quarter closed.
I cut the scope to the one loop that proves the thesis: upload documents, ask a question, get a cited answer. Everything else — teams, billing tiers, integrations — went on a list for after the evidence came in.
The build itself ran on an AI-accelerated workflow: generated scaffolding, auth, upload handling and test fixtures in the first days, which left the calendar free for the parts that actually decide whether a RAG product feels good — chunking strategy, retrieval quality, streaming and citation UX.
Most bad answers were bad retrieval, not bad generation. Tuning chunking and adding a rerank step moved accuracy far more than swapping models did.
Every claim links back to the source passage. It turned "interesting demo" into something the team was willing to let customers rely on.
Answers stream from the first token at ~1.4s, so a five-second answer feels immediate instead of broken.
Cutting the roadmap down to upload → ask → cited answer, and writing down what we were deliberately not building.
Auth, uploads, database and the app shell generated and reviewed inside the first sprint.
Chunking, embeddings, reranking and an eval set to stop quality regressing silently.
Production deploy, cost and latency budgets, and event tracking on every question asked.
Measured across the first month of real usage with early customers.