AI Lakaton: Hackathon Compliance Auditor

AI Lakaton is a compliance auditor built for hackathon organizers and participants. You provide the official rules (pasted text, a PDF, or a URL) and the project to check (a public GitHub repo or a ZIP upload), and the app returns a full compliance report: an overall score, a requirement-by-requirement checklist, cited evidence for every claim, warnings, and suggested fixes — all exportable as a Markdown report. The core design principle is that it never invents evidence. Under the hood, a 3-agent LLM pipeline splits the work: a Rules Agent extracts every concrete, checkable requirement from the rules text without seeing the project; a Project Audit Agent evaluates each requirement against the submitted project and must cite a direct quote or file path as proof; and a Review Agent can only make verdicts stricter, never looser. Every "Complies" verdict then goes through a deterministic anti-hallucination cross-check in plain Python: if the cited evidence doesn't literally appear in the project text, the claim is automatically downgraded. The final score is also recomputed from fixed per-status weights instead of trusting the model's own number, so the result stays grounded and reproducible even when the AI is wrong. If no AI provider key is configured, the app falls back to a fully deterministic keyword-matching mode, so it still works offline or without API credits. Built with a FastAPI/Python backend and a Next.js/React/TypeScript frontend, deployed as a single Vercel project.