Label the work before you choose the model
Most stalled ML projects share a quiet problem: labels were treated as a side task. Here is how we pressure-test labelling capacity early.
Field notes
Writing from the floor of implementation work.
Articles stay close to labelling practice, evaluation honesty, and ownership — the topics that decide whether an ML feature survives contact with production.
Offline metrics that survive a sceptical review
Accuracy alone rarely persuades a risk committee. Pair overall scores with slice reports and a stated cost of false positives.
When not to automate with machine learning
Some decisions look like prediction problems until you examine volume, latency, and the cost of being wrong.
Handing models to on-call without drama
Production ML fails quietly when nobody owns drift alerts. Define ownership before the first night-time page.