Front-end developer turned AI-assisted audit specialist.
I build methodologies for auditing legacy codebases with adversarial AI agent teams using Claude Code.
legacy-audit-agents — An open-source methodology for auditing legacy codebases with dual-team adversarial AI agents.
Built from 7 real iterations on a 250K LOC, 20-year-old monolith:
Iteration 1: 1 agent → hallucinated 4 classes that didn't exist
Iteration 7: 11 agents → 81.8% reliability score, adversarial-verified
168 verified findings. 889 "foreign keys" corrected to 15 real. 4 AI hallucinations caught and eliminated. Every mistake documented so others don't repeat them.
The repo includes progressive setups (Level 2→7) so you can iterate the same way — not just copy the final config.
TypeScript · React · Vue/Nuxt · Tailwind · Python/FastAPI · Claude Code · Agent Teams
- Proof over assumptions — every claim needs
file:lineevidence - Adversarial review — if a second team can't disprove it, it's probably true
- Documented failure — publishing what went wrong is more valuable than showing what worked
