rahuldr556/ast-guided-llm-code-review_system

This project is a code review tool that automatically scans Python programs for bugs, combining two techniques: structural code analysis (called AST, which reads how code is organized like a grammar checker reads sentences) and an AI language model that reasons about what the code is doing. The goal is to catch real bugs before humans have to find them manually. The system works as a five-step assembly line. First it parses the source file to understand its structure. Then it runs detectors that look for specific problem patterns. Next it slices out the relevant code sections around any suspicious spot so the AI does not have to read everything at once. The AI then reviews those slices and judges whether each flag is a real bug. Finally an aggregator collects all findings and resolves any conflicts between detectors. The tool looks for six specific types of bugs: accessing a variable that might be None, dividing by zero, reading outside array bounds, data leakage between training and test sets in machine learning code, tensor shape mismatches, and resource handles that are opened but never closed. These cover common mistakes in both ordinary Python code and data science pipelines. Benchmark results reported in the README are strong. On a synthetic test set of 80 cases the tool scored 100% precision and 94.3% recall, meaning it flagged only real bugs and caught almost all of them. On 12 hand-picked real-world bugs the recall was 92%. When run against 90 real-world cases using GPT-4 the recall was 88%, still with perfect precision. You can run it in simulation mode at no cost, which uses a built-in reasoning layer without calling any external API, or with an OpenAI API key to use GPT-4 for the review stage. The repository includes 39 unit tests and 92 benchmark cases across 25 Python files totaling roughly 8,500 lines of code.