AI Project Reviewer.
A Java application that automatically evaluates a software project's source code using deterministic static analysis, a sandboxed build check, and real LLM calls, then generates a LaTeX evaluation report.
PROBLEM
Manual code evaluation is time-consuming and inconsistent. Existing tools often lack the ability to execute untrusted code safely or integrate multiple LLM providers without vendor lock-in.
DATA
The system analyzes project source code by exploring file trees, running static analysis rules, and executing a sandboxed build check to validate the project manifest (pom.xml/package.json).
APPROACH
Built with a focus on architecture: low coupling, justified design patterns, and LLM interchangeability. The system uses a ResilientLLMProvider wrapper with retry/timeout/fallback logic, allowing seamless switching between Mistral, Groq, and local models. Docker is used for secure, network-isolated code execution.
RESULT
Successfully generates .tex and .pdf evaluation reports with a history of past runs. The project demonstrates robust software architecture, safe handling of untrusted code, and resilience to LLM failures.
