Role: AI Engineer, LLM featuresFull interview loop

For a team with a model-backed feature in front of customers, or about to be. It assumes the candidate will own that feature rather than research a new one, so there is no round on model architecture and none on training.

The four rounds are deliberately the same system seen four ways. The first is the code immediately around one model call and what it does when the call fails. The second is what may be reused across users and what a wrong reuse costs. The third is whether any of the numbers being reported about the feature are true. The fourth is what the system is allowed to do when the text it read was written by someone hostile.

LLM systems design is read three times because it is the competency the role turns on and the one most easily faked in a single conversation: once in code, once as a design argument, once as a question about capability. Evaluation and production judgment are read twice each. Coding is a nice-to-have here rather than a must, which is a real trade — this loop will pass a candidate who writes ordinary code and thinks clearly about failure over one who writes elegant code and has never measured a model's output. 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, and it is scored against an incident the candidate's own code caused. That is the only round of the four with a wrong answer that is not technical.