ProofStack: building verification for AI-proposed code fixes

Nebius x NVIDIA AI Hackathon prototype

The problem • AI tools suggest security fixes that look right but can break the build or something else

The idea • The model only proposes a patch • Deterministic checks should decide whether it is accepted

Intended workflow Step 1 - Prove the vulnerability with a probe or test Step 2 - NVIDIA Nemotron (via Nebius) proposes a scoped patch Step 3 - Verification gates check it: syntax, lint, security re-test, regression, concurrency, performance, resources, patch hygiene Step 4 - The result is recorded in a SHA-256 hash-chained ledger

What I built • Sandbox isolation • Benchmark testing • Verification gates • Evidence ledger with a tamper simulator

Where it stands • The verification gates are implemented as standalone modules • End-to-end orchestration is still demo-stage • The hosted demo and screenshots below use sample data • Not client work

Stack: Next.js, FastAPI, Python, Docker, NVIDIA Nemotron

Demo: https://proofstack-app.vercel.app/ Code: https://github.com/nikhillakra2007-tech/ProofStack 2-min video: https://youtu.be/EsdJ2vkg5HA

Open to remote full-stack and AI-integration work (Next.js, FastAPI, Python).