We made the same mistakes most teams make. We focused on the model. We got the signal to noise wrong. We underestimated what it actually takes to make review guidelines hold across a large codebase.

This talk is about what we learned fixing those mistakes. We will cover context construction: why we started with static embeddings, where they broke down, and what we traded off moving to dynamic RAG and hybrid approaches. We will cover personalization at scale: how to maintain meaningful review standards from a single team up to org-wide without them collapsing into noise. And we will cover the part that changed everything: building an evaluation platform and AI-driven feedback pipeline that captures developer feedback, diagnoses where the system falls short, and ships fixes fast.

The core insight is that the model is not the product. The system around it is. You will leave with a clear framework for building AI developer tools that get better over time, and a sharper sense of where most teams are still getting it wrong.