The detector that finally met my launch bar is a logistic regression over 40 handcrafted features. It flagged 0 of 100 fresh human essays and caught 150 of 150 AI essays on a preregistered, frozen benchmark. The interesting part is the two failed rounds that made it work.
I took a production FastAPI backend from 334 CRAP violations to zero in a day. The interesting part isn't the number — it's how I proved the refactor didn't change behavior, and the 13 drifts an independent diff review caught that the tests didn't.
I pointed a team of AI agents at ~300 of my own repos and asked whether I actually use coding agents the way the 2026 practitioners preach. The first scorecard flattered me. A rival model caught it lying. Here's the honest verdict and what I rebuilt in response.
How we built an AI-powered C++ code auditor that checks four pillars — memory safety, performance, modern C++ idioms, and architecture — by distilling the C++ Core Guidelines, Abseil, Scott Meyers, and Fedor Pikus into a single CLI tool.
How I built an AI-powered .NET performance advisor agent by distilling knowledge from David Fowler, Stephen Toub, and Microsoft's official docs into a single Claude Code agent.