AI Code Review Assistant

Problem

Modern code review is the bottleneck of shipping. Human reviewers are scarce, and the most common feedback is mechanical: “add a test”, “misleading variable name”, “duplicate of function above”.

Solution

• Built a GitHub App that runs on every PR, indexes changed files into a context store, and asks a fine-tuned LLM for line-anchored comments.
• Designed prompts so every suggestion is grounded in a specific file and line range, with refusal when confidence is low.
• Integrated with CI: blocks merge only for high-signal checks, otherwise posts soft comments.

Result

• Adopted by 40+ open-source projects in three months, median time-to-first-review of 35 seconds.
• Reviewer survey: 72% said the bot made reviews shorter and more focused.
• Fully self-funded through GitHub Sponsors.