Role: Machine Learning Engineer, production modelsFull interview loop

For a team with a model making decisions in production, or about to be, and an engineer who will own it end to end rather than hand a notebook to someone else. It assumes the candidate will train models as well as ship them, which is the difference between this loop and the AI Engineer one: there is a round on the modelling decision and a round on retraining, and there is no round on prompting or on LLM features.

The four rounds are the same model seen at four moments in its life. The first is the code that computes its features, and the ways the training copy and the serving copy of that code disagree. The second is the decision about what to train at all, made from a data pack rather than from habit. The third is the experiment that says the new model is better offline and worse online, and what to believe. The fourth is the retrain that quietly moved every score and cost money before anyone noticed.

Production ML systems is read three times because it is the competency the role turns on and the one a single conversation is least able to tell apart from having read about it: once in code, once in an experiment readout, once as a design. Modelling judgment and experimentation are read twice each, and each second read comes from a different direction — the modelling call is made prospectively in the second round and diagnosed retrospectively in the third; the experiment is designed in the second and decided in the third. Coding is a nice-to-have here rather than a must, which is a real trade: this loop will pass a candidate who writes plain code and reasons carefully about where two systems disagree over one who writes elegant code and has never measured a model against production. If that is the wrong trade for your team, add a general coding round and make it a must.

Ownership is scored once, in the fourth round, against an incident the candidate's own retraining job caused. It is the only round of the four whose wrong answer is not technical.