Phoenix
Agentic generative AI for safe codebase refactoring with strict safety gates
Image will load when scrolled into view
LLM AgentsSafety GatesEvaluationStatic Analysis
About the Project
Phoenix is an agentic system for safe end-to-end codebase refactoring. It pairs LLM-driven patch synthesis with strict static-analysis and dynamic-validation safety gates, so that no generated change lands without passing correctness checks.
I architected the agentic pipeline and designed a multi-repository evaluation protocol quantifying correctness preservation, maintainability gains, and developer-effort reduction.
The work is published as 'Phoenix: Safe End to End Codebase Refactoring via Multi-Agent LLMs' at the International Conference on Responsible AI (ICRAI) 2026.
Project Details
StatusPaper at ICRAI 2026
Role
Co-author & Architect
Stack
LLM Agents
Static Analysis Gates
Dynamic Validation
Python