ProofStack - AI Code Repair Verification (Hackathon Prototype)

ProofStack is a prototype I built with a teammate for the Nebius x NVIDIA Global AI Hackathon. It explores one idea: AI tools suggest security fixes that look right but may not compile or may break something else, so the model should only propose a patch and deterministic checks should decide whether it is accepted. How it is meant to work: 1) prove the vulnerability with a test or probe, 2) have NVIDIA Nemotron propose a scoped patch, 3) run the patch through verification gates (syntax, lint, security re-test, regression, concurrency, performance, resources, patch hygiene), 4) record the result in a tamper-evident SHA-256 hash-chained ledger. What the project includes: 19 red/green vulnerability probes with fixtures, standalone verification gates, 12 DAST web-security checks, container-based sandbox runners, a Nemotron model client, a SHA-256 evidence ledger with a tamper simulator, and a Next.js and FastAPI cockpit with a side-by-side diff view. Honest status: this is a hackathon prototype, not client work. The hosted demo uses scripted sample data for the container and AI-fix steps, and the screenshot above is sample data. The full pipeline needs a local setup and a Nebius API key, and the backend does not yet run every gate end to end. Stack: Next.js, FastAPI, Python, NVIDIA Nemotron (via Nebius), Docker, SQLite/PostgreSQL. Live demo: https://proofstack-app.vercel.app/ Source: https://github.com/nikhillakra2007-tech/ProofStack 2-minute video: https://youtu.be/EsdJ2vkg5HA