- TL;DRwhat you say if they cut you off
- ML CI/CD retains everything standard CI/CD does (lint, tests, build, deploy) and layers on three items, because the model is a function of data, not code alone: validate incoming data (schema, distribution), gate on model quality (a retrained model has to beat the incumbent on a held-out, time-correct eval, not just clear unit tests), and version data + features + model + code together so any run reproduces. Deployme…
- HOW TO APPROACH ITthe order, and the clarifying questions
- Note the shared base in one sentence, then spend your time on what is ML-specific. The framing that scores: in software, behavior comes from code; in ML it comes from code and data, so the pipeline must test, gate, and version the data path as well. The rest follows from that.
- THE FOLLOW-UPwhat they probe next, with the strong reply
- "Why isn't passing unit tests enough?" Unit tests check code correctness; they cannot tell you the model got worse. You need an eval-vs-baseline gate.
- THE MISTAKEwhat sounds right and loses the offer
- Describing plain software CI/CD and missing data validation and the model-quality gate. Promoting a model because the pipeline "passed" without comparing it to the incumbent. Versioning only code, so a production model cannot be reproduced or audited. Straight-to-prod deploys with no canary, trustin…