Patchworks

Inspiration

AI coding tools can generate code quickly, but generating a patch is only one part of fixing a bug. Developers still need to understand the issue, inspect the proposed changes, run tests, and decide whether the fix is safe to merge.

We built Patchworks to bring that entire process into one workbench, with a focus on making AI-generated repairs inspectable and verifiable.

What it does

Patchworks is an autonomous code-repair workbench that analyzes repositories, identifies potential bugs and security issues, explains findings with code evidence, and generates targeted patches.

Developers can inspect unified diffs, run verification tests, repair multiple findings in batches, track related GitHub pull requests, and prevent duplicate repair submissions. Failed repairs can be rolled back, and pull-request publication remains under developer control.

How we built it

The backend uses Python and FastAPI to orchestrate repository inspection, model inference, patch generation, verification, and GitHub integration. The frontend is built with React and TypeScript, providing a dark, interactive workbench for findings, code diffs, repair actions, and verification results.

We integrated Nebius Token Factory for model inference using NVIDIA's Nemotron model. We also implemented repository-aware pull-request tracking, duplicate-publication safeguards, and batch repair workflows with per-finding verification.

Challenges we faced

One of the main challenges was making the repair workflow more than a code-generation interface. Generated changes need to be isolated, tested, and reported honestly. Batch repairs add further complexity because a failed repair should not erase previously verified work.

We also worked on recognizing existing pull requests so the system can distinguish a new repair opportunity from a fix that may already have been submitted.

What we learned

Building Patchworks reinforced that useful coding agents need more than capable models. They need reliable orchestration, clear evidence, safe change management, meaningful verification, and human oversight.

We tested the workbench against its own repository, using its findings as a starting point for further investigation. This gave us a practical way to examine the tool's output on a real codebase while recognizing that model-generated findings still require validation.

Our goal is to make AI-assisted code repair more transparent, testable, and accountable.

Built With

• fastapi
• git
• github-api
• nebius-token-factory
• nvidia-nemotron
• pytest
• python
• react
• render
• rest-api
• scikit-learn
• supabase
• typescript
• vite