DocuMind
An AI support agent that answers questions over a real, messy technical knowledge base (a product's docs + GitHub issues + Slack export) and knows when it doesn't know.
Build a retrieval-augmented support agent over a real, messy knowledge base, then make it measurably better and operationally trustworthy. The through-line is honesty: a system that knows when it does not know, and an owner who can prove it does.
Foundation: get a working RAG pipeline end to end
Ask a real question about your knowledge base and get an answer with accurate, verifiable citations - running locally, end to end.
Applied: evals, error analysis, and measurable improvement
Stop guessing whether changes help. Build the measurement apparatus first, then improve something and prove the improvement.
Advanced: agentic workflow, guardrails, and production monitoring
Handle multi-step tool-using questions, reject bad inputs and outputs, and run monitoring that would catch a regression without you manually testing it.