CodeEnforcer

CodeEnforcer is an AI agent that automates GitHub Pull Request code review using large language models running on AMD Instinct MI300X GPU hardware. Every software engineering team spends significant time on manual code reviews. Senior developers read through hundreds of lines of changed code, identify bugs, flag security vulnerabilities, and write feedback. This process is slow, expensive, and inconsistent across reviewers. CodeEnforcer solves this by automating the entire process using AI. When a user pastes a GitHub PR URL into the CodeEnforcer web interface, the agent fetches the complete code diff using the GitHub API. The diff is then split into optimized chunks that the language model can process efficiently. Each chunk is sent to DeepSeek Coder 1.3B — a code-specialized large language model — served via vLLM on AMD Instinct MI300X GPU using the ROCm 7.2 software stack. The model analyzes the code changes and returns structured JSON containing every issue found, with severity classification, line references, problem descriptions, and specific fix suggestions. All findings are aggregated, sorted by severity, and displayed in a clean dark-themed web interface. The AMD MI300X GPU with 192GB VRAM enables fast, consistent inference at 90.9 tokens per second with an average inference time of 2.20 seconds per chunk. Benchmark results across 5 runs showed near-zero variance, demonstrating the reliability and performance of the AMD hardware stack. CodeEnforcer is built entirely on open source tools — Python, FastAPI, PyGithub, DeepSeek Coder, vLLM, and ROCm — making it fully reproducible and extensible. The complete source code is available on GitHub with setup instructions and benchmark results.