Short:

Most AI code review tools impress in demos and disappoint in production. This talk goes under the hood dissecting how LinearB's AI reviews over 5K pull requests daily with a 99.4% satisfaction rate. Hint: it wasn't done by nailing the model on day one, but by building a system that learns from every review.
We review what makes an AI code review system genuinely useful; our architecture choices for building context (e.g. dynamic RAG) and prompt engineering; and the evaluation system; and at the core - the AI driven feedback loop that powers continuous improvement.

Longer:

Most AI code review tools impress in demos and disappoint in production. LinearB AI reviews over 5K Pull requests daily with a 99.4% satisfaction rate - not because we nailed the model on day one, but because we built a system that gets smarter with every review.

This talk goes under the hood of LinearB AI Code Review. We'll cover what makes an AI code review system genuinely useful (hint: it's not just signal to noise ratios). We'll walk through context preparation strategies such as static embeddings, dynamic RAG, and hybrid approaches - with honest tradeoffs and why we made the choices we did. And we'll show how we elegantly support personalization from team practices up to org-wide standards.

But the core of the talk is the improvement loop: how we built an evaluation platform and AI-driven feedback pipeline that captures developer feedback, diagnoses where the system falls short, and ships fixes fast. Using AI to improve an AI tool is where things start to seem magical.

You'll leave with a practical framework for building AI developer tools that compound, turning every interaction into a step toward better.

Key Takeaways:

- What distinguishes AI Code Review systems developers embrace from ones they mute
- How to choose the right methods for context construction
- The evaluation + feedback AI architecture that powers continuous improvement at scale

Best For: Engineering leaders evaluating AI dev tools, platform engineers building them, and practitioners who want to understand what separates AI features that stick from those that get turned off.