Architecture deep-dives, research notes, and the occasional opinion — written by the engineers building deterministic code review at codingassist.bot.
Why we built a fourth thing — and what makes it different from copilots, static checks, and human reviewers.
A tour of the reasoning model that powers the codingassist.bot pipeline, told through one example PR.
Diff + codebase graph + ticket contract. Why fusing these three signals before reasoning eliminates the entire class of "plausible-but-wrong" reviews.
An argument for treating PR comment language as a real engineering interface, not a social one.
A dataset, a methodology, and three things we got wrong before getting one thing right about reviewer attention.
Replaying everything is a quadratic problem. We picked a budget and let it shape the system. The constraint became the design.
New posts, behind-the-scenes notes, and the rare paper we couldn't put down. No marketing emails, no growth-hack subject lines.