Bubo: AI code review that learns from maintainer feedback

I built Bubo because I’m tired of AI reviewers flooding PRs with false positives, then learning nothing when a developer explains why a finding is wrong. I started with three requirements: simple setup, evidence-backed findings or LGTM, and learning from human comments on findings so the reviewer becomes tuned to each repository. I ran a small directional benchmark on 20 pinned PRs/MRs using GPT-5.5: - Bubo: 7/8 defects, 27 findings, 0% noise - ai-codereviewer: 6/8 defects, 118 findings, 20% noise - ChatGPT-CodeReview: 5/8 defects, 75 findings, 11% noise It is a small sample and I picked the PRs, so I treat it as encouraging rather than definitive. The harness and PR list are in the repository. Current limits: Python 3.14+ and polling rather than webhooks. My next direction is pluggable subject-matter specialist Skills instead of one general reviewer. That routing is roadmap, not shipped today. Bubo is currently running in production in two places: a large data-processing/ETL codebase and a fintech crypto stack. I’d appreciate technical feedback on whether the evidence-or-LGTM and repository-learning approach holds up.