A staff engineer shares how he built an AI code reviewer trained on three years of his team's PR comment history. Starting from a generic checklist, the system evolved to mine GitHub PR comments via API, generate per-engineer reviewer profiles using Claude, and apply them all in a single pre-review pass. A multi-agent V2 ran parallel reviewer agents but cost $20 per PR. V2.1 reduced cost ~7x to $3 by adding a deterministic file classifier to strip irrelevant profile sections, using cheaper Haiku/Sonnet models for most passes, and reserving Opus only as a final skeptic agent to prune false positives. The open-source template repo is available to clone and adapt in an afternoon.
Table of contents
V0: A generic checklistV1: Let’s mine the historyV2: We got greedyV2.1: Tiered intelligence3 takeaways for your own implementation:Discover weekly27.2K Impressions