Sentry's AI Code Review tool integrates production data to predict bugs before code is merged. Part of Sentry’s AI debugger, Seer, it uses historical data and context from the application to identify real issues instead of generating irrelevant feedback. The system aims to improve the reliability of code reviews by focusing on potential bugs in the specific changes being made, rather than overwhelming developers with false positives. The architecture of the code review system includes a multi-step pipeline that filters pull request (PR) data, predicts bugs, and packages suggestions for review. Initially, it filters PRs to focus on the most error-prone files, especially in large changes. The prediction phase involves generating hypotheses about potential bugs using various data sources, including the actual code changes, PR descriptions, and historical Sentry data. The system learns from each PR, accumulating "memories" that help refine its predictions over time. An example from a specific commit illustrates the system in action. The analysis identified missing error handling in analytics calls, which could lead to user-facing errors if exceptions propagate. It also flagged a potential issue with a missing `user_id` parameter in a refactored analytics call, raising questions about whether this would affect data collection. The detailed feedback provided by the AI enhances the overall quality of code by addressing these critical areas before they can lead to bugs in production.