When a simple algorithm beats ML
Draft- engineering
- tooling
The brief was simple: stop PRs from sitting unreviewed because nobody knew who should look at them. The obvious move in 2024 was to throw a model at it. I didn't.
Draft — full write-up coming. Outline below.
- The problem: reviewer assignment as a routing problem, not a prediction problem.
- Why the git history already contains the answer (ownership + recency).
- Exponential-decay weighting, and why recency should dominate.
- The part that mattered: a recommendation an engineer can audit beats a better one they can't.
- Where ML would actually have helped — and why it wasn't here.