A newer loop with less settled conventions. Expect evaluation to dominate — how you measure whether a non-deterministic system is working — alongside retrieval design, prompt strategy and cost per request.
Also advertised as AI Engineer, LLM Engineer, GenAI Engineer — the loop is the same, and so is the practice below.
Data loops test two things that pull against each other: rigour with methods, and the judgement to know when a rough answer is the right one. Candidates who are strong on one are often visibly weak on the other.
Expect questions where the data is deliberately insufficient. Stating the assumption you are making, and what would change your answer, is most of the mark.
Each round is run by the interviewer built for it, with its own scoring axes — not one generic interviewer asked to change subject.
Metrics, experiments, causal reasoning, and decisions under uncertainty.
Scored on Analytical communication · Data & inference rigor · Decision quality
Metrics, experiments, causal reasoning, and decisions under uncertainty.
Scored on Analytical communication · Data & inference rigor · Decision quality
Hands-on problem solving, code quality, debugging, and engineering judgment.
Scored on Problem-solving communication · Engineering correctness · Adaptability & craft
System design, trade-offs, and war stories from production.
Scored on Technical communication · Engineering depth · Ownership & collaboration
Metrics, experiments, causal reasoning, and decisions under uncertainty.
Scored on Analytical communication · Data & inference rigor · Decision quality
Six you can expect, in the register interviewers actually use. Answer them out loud before you read the next section — reading a question and answering one are different skills, and only the second is marked.
The gap between these two is most of your score, and it is easier to see than to be told.
“How do you evaluate a system whose output is different every time?”
"We check the outputs look good." Manual spot-checking does not scale past a handful of cases and cannot catch a regression.
A held-out set with graded cases, a scoring method you can defend — deterministic checks where possible, model-graded where not, human-labelled on a sample — run on every prompt or model change. Evaluation is the discipline of this role and the loop is mostly about it.
The post-mortem nobody sends you. These are specific to this loop rather than general interview advice.
For AI / LLM Engineer, a general round runs with Marcus — metrics, experiments, causal reasoning, and decisions under uncertainty. It is the fastest way to find out which round you actually need to work on. The first session is free.
Practise a AI / LLM Engineer interview →Before you practise, it is worth reading how the common rounds are marked: tell me about yourself, behavioral questions and STAR, and system design.