Inspiration

Traditional static linters like Flake8, Pylint, and SonarQube flood developers with noisy formatting warnings without answering what matters in production and hiring: What is the probability this code will fail? and Does this developer write resilient systems? Competing in Track 1: Future Tech (FT-05: AI-Powered Automated Code Review) at Neural Storm 2026, we built ProofCommit Pro to unite automated code review with verifiable developer skill evidence inspired by sponsor 9AM (by TechMojo).

What it does

ProofCommit Pro (https://proofcommit-pro.vercel.app/) is an AI & ML code auditing platform that turns Python commits into defect predictions, git patches, and verifiable skill certificates: 8-Dimensional Static AST Profiler: Extracts Lines of Code, AST Node Count, Cyclomatic Complexity, Max Nesting Depth, Function Count, Exception Handlers (try/except), Assertions, and Docstring Coverage. Random Forest Defect Classifier: Feeds the 8 AST metrics into a scikit-learn Random Forest model (n_estimators=100, max_depth=6) to predict 0–100% Defect Risk across Low, Moderate, and High risk tiers. Gemini AI Patch Synthesis: Generates severity-ranked findings (High, Medium, Low, Info) and a copy-pasteable unified git diff (+ / -) patch using Google Gemini (gemini-2.5-flash). 9AM Skill Evidence Scorecard & Razorpay Checkout: Converts code quality into a 0–100 Skill Score (9AM Gold / Silver / Bronze) and unlocks a downloadable .pdf / .png 9AM Verified Skill Certificate via Razorpay (₹299 INR) or 1-click Judge Demo Mode.

How we built it

Built with Python (ast, scikit-learn, NumPy, FastAPI), React 18 (Vite), Tailwind CSS, Google Gemini API, Razorpay, and deployed on Vercel with an in-browser serverless AST/ML engine for zero cold-start latency. On a 100-sample stratified holdout test set, our Random Forest model achieved: Accuracy: 89.2% | Precision: 87.5% | Recall: 85.4% | F1-Score: 86.4% Top Features: Cyclomatic Complexity (38.4%), Max Nesting Depth (27.2%), Exception Handlers (18.6%), and Docstring Ratio (15.9%).

Challenges we ran into

Early heuristics penalized longer files simply for having more branches; normalizing complexity per function and weighting defensive try/except blocks and docstrings fixed scoring accuracy. Combining deterministic AST metrics with ML classification and LLM patches prevents hallucinations on line counts and complexity figures. AI Tool Disclosure (Neural Storm Rules): We used Google Antigravity, Gemini, Claude Sonnet, and Cursor for UI scaffolding and rapid prototyping while directing the AST schema, Random Forest pipeline, and Razorpay integration.

What's next for Proof commit.pro

GitHub App CI/CD webhook integration for automated PR comments and multi-language Tree-sitter AST support for TypeScript, Go, and Rust.

Built With

Share this project:

Updates

Submission history