Mini-Jarvis Onboarder

Every new engineer joining a codebase loses their first week just figuring out how things work and it costs a senior teammate real time to walk them through it. Mini-Jarvis Onboarder fixes this with a five-agent AI pipeline: a Repo Analyst maps the codebase structure, a Docs Agent extracts setup steps from the README, and three Gemma-powered agents generate a plain-language architecture briefing, 3-5 starter tickets grounded in the actual file tree (not invented files), and a personalized welcome message all in about a minute. The architecture follows a deliberate safety pattern: agents are stateless, the orchestrator runs a fixed deterministic sequence rather than letting an LLM decide what to do next, and every output is a draft nothing is sent until a human reviews and confirms it. Built with FastAPI, Gemma via Fireworks AI on AMD infrastructure, and Docker. A live on-demand Gemma deployment was configured and validated end-to-end during development; the submitted demo runs with a documented mock-mode fallback due to shared GPU capacity constraints during the hackathon window, exercising the identical pipeline logic, JSON hallucination mitigation, and retry handling.