RepoMedic AI: An AI Agent That Fixes Bugs—and Proves the Fix Works

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

## What I Built: RepoMedic AI

I built RepoMedic AI, an AI-powered bug repair agent that helps developers find bugs, generate code fixes, and verify whether those fixes actually work.

Traditional AI coding assistants can suggest solutions, but generated code is not always correct and may introduce new bugs. RepoMedic AI addresses this problem with an evidence-gated repair workflow: it analyzes a bug, proposes a patch, checks the changes against safety rules, runs eligible tests in an isolated environment, and presents the results so developers can inspect the evidence before trusting a fix.

The project is designed for software developers, computer science students, open-source contributors, freelancers, and engineering teams that want to spend less time debugging and more time building reliable software.

Who would pay for it? Freelance developers and software agencies could pay to reduce debugging time, while startups and engineering teams could use it to speed up code reviews and routine bug fixing. Open-source maintainers could benefit from faster issue investigation, and educational institutions could use it as a practical tool for teaching debugging and software testing.

The goal is simple: make AI-assisted debugging more transparent, secure, and dependable. Rather than accepting an AI's claim that a bug is fixed, developers get a proposed patch and test evidence they can evaluate themselves.

Explore RepoMedic AI:

• Live demo: https://repomedic-ai-stqt.onrender.com/
• GitHub repository: https://github.com/MDADILIFTEKHAR/RepoMedic-AI

RepoMedic AI is currently a working project and public demo, with opportunities to expand into paid developer tools and team-focused plans as the product matures.

Demo

repomedic-ai-stqt.onrender.com

Code

MDADILIFTEKHAR / RepoMedic-AI

RepoMedic AI 🩺

RepoMedic AI is an open-weight AI-powered Python bug-repair harness built around an Evidence-Gated Repair Loop.

Unlike conversational coding assistants that rely on model self-attestation ("I fixed the bug!"), RepoMedic AI enforces strict, deterministic verification: a proposed unified-diff patch is never considered successful until it passes deterministic policy gating and executes cleanly in an isolated test environment with zero regressions.

---

🎯 Problem Statement

Modern coding models frequently hallucinate that a patch works, propose sprawling changes that touch unrelated files, tamper with test assertions to make failing suites artificially pass, or generate insecure code. Furthermore, executing arbitrary untrusted code directly on developer machines poses significant security risks.

RepoMedic AI solves this by introducing:

1. Deterministic Security & Policy Gating: Unified diffs are inspected before application. Path traversal (../), changes to protected files (requirements.txt, .git/, CI workflows), and test alteration tampering are unconditionally blocked.
2. …

View on GitHub

## How I Built It

I built RepoMedic AI around an open-weight coding model, Qwen2.5-Coder, with a custom Python-based agent harness designed to do more than generate code. My goal was to build an agent that can propose a bug fix, evaluate the proposed changes, run tests, learn from failure evidence, and report whether the repair actually passes verification.

1. Open-Weight AI and Local Inference

The core model is Qwen2.5-Coder:7B, an open-weight model specialized in code generation, code reasoning, and code repair. I integrated it with Ollama, which allows developers to run the model locally without depending on a hosted inference API.

For hosted inference, RepoMedic AI also supports configurable OpenAI-compatible endpoints, including providers such as OpenRouter, Together, and self-hosted vLLM deployments. This gives developers flexibility to choose between local inference and hosted model access.

2. Building the Agent Around Evidence

Instead of relying on a prompt that simply asks an AI to fix a bug, I built a custom, evidence-gated repair loop in Python.

The workflow begins by scanning the target repository and running baseline tests to identify existing failures. The agent sends the relevant context and failure evidence to the coding model, which proposes a repair as a unified diff.

Before applying the patch, a deterministic policy gate checks whether the changes comply with the project's safety rules. It rejects prohibited modifications, including test tampering, path traversal, and changes to protected files.

Accepted changes are evaluated in a disposable Git worktree or temporary workspace, keeping experiments separate from the original repository. Tests can run in restricted Docker containers with networking disabled and resource limits. If a test fails, RepoMedic AI extracts the actual failure output and feeds it back into the model for another repair attempt.

I limited the repair process to three attempts to keep the workflow bounded and predictable. A patch is considered verified only when it passes the required policy checks and test execution.

3. Developer Experience and Evaluation

I built a web-based developer interface with live repair progress, syntax-highlighted diffs, test evidence, and benchmark results. I also added a command-line interface and REST API so the repair workflow can be used programmatically.

To evaluate the approach, I created a deterministic benchmark containing 10 Python repair tasks. In the recorded results, one-shot repair achieved a 20% pass rate, while the bounded repair loop achieved a 50% pass rate. This is a 30-percentage-point improvement on that small benchmark, not a guarantee of real-world repair accuracy.

The project also includes an Agent Skill file following the Agent Skill Open Standard, helping make the repair capability discoverable by compatible agent environments.

4. Deployment and Open Innovation

I deployed the web application on Render and published the source code on GitHub. In hosted demo mode, users can explore the predefined benchmark fixtures without enabling unrestricted execution of arbitrary repositories on the public server.

Open-weight AI matters to this project because developers can run Qwen2.5-Coder locally, experiment with their inference setup, and maintain more control over their code and model environment. The interchangeable provider configuration also makes it possible to adapt the system to different infrastructure needs.

My main design principle is simple: the model proposes the fix, but deterministic checks and test results decide whether it works.

Project links

Live demo: https://repomedic-ai-stqt.onrender.com/

GitHub repository: https://github.com/MDADILIFTEKHAR/RepoMedic-AI

Why Does Open Innovation Matter?

My Agent Session

My First User

For your RepoMedic AI submission, remove the “Prize Categories” section. The AWS Builder Center Weekend Challenge requires the #agents tag and the specified article content; it does not list separate partner prize categories.

AWS Builder Center

Note: Render is your project's deployment platform, not a prize category.
Make sure #agents is added as the article tag.