AI Code Review: Comparing Ready-Made Tools and Building Your Own
Should your team adopt an existing AI code reviewer or build its own? I will compare both approaches using real development workflows and moreover, tell the most effective approaches in verifying the PRs by AI reviewers. We will examine CodeRabbit, Greptile, Cursor Bugbot, and Claude Code across review quality, false positives, repository context, customization, developer experience, privacy, maintenance, and cost. Using PRs with planted defects, we will see what these tools catch and miss. Next, I will show how we built DevDigest, our own AI code reviewer. We will explore its architecture, model integration, structured findings, and context pipeline: system instructions, PR intent, linked specifications and skills, a repository map, affected callers, and finally the diff. We will also integrate review with spec-driven development. Instead of checking only whether the code looks correct, the reviewer can validate whether the implementation follows the original requirements, architectural decisions, and expected behavior. The talk will cover the pros and cons of ready-made and custom solutions and caveats on our path of creating a well-functioning solution that works for our purposes. I'm also planning to integrate Antigravity and Firebase into the talk. The duration is flexible, from 30 to 60 min. 30min can be also enough.
Speakers: Ivan Lapa
Explore the event schedule.