Role: Forward Deployed Engineer, AI deploymentsFull interview loop

For a team hiring an engineer who will spend most of their time inside customers' organisations, deploying model-backed systems into estates they do not control. It assumes the candidate will own an engagement from discovery to handback, so there is no sales or demo round, and nothing on training models.

The first two rounds read the same two competencies in two modes. Round one puts the candidate in a new engagement, live, with an interviewer playing a customer who volunteers nothing; round two reconstructs a real engagement the candidate chose. Discovery and customer judgment are each read once in a room the candidate does not control and once in a story they do, which is the pairing the method asks for and the one most likely to separate an engineer who has run engagements from one who has watched them. Deploying into customer systems is read once in conversation and once in code: round three hands the candidate an export from a system they cannot change, with a security review waiting behind it.

The simulation comes first, on purpose. It scores what a candidate thinks to ask, and in what order, before anyone has told them what matters, and every later round asks about something its customer withholds: the first week on site, who gets access and when, how a legacy system's files really behave, what data may leave. A candidate who had just been asked about those would no longer show what the simulation reads. The rounds after it score what the candidate did or wrote, which hearing the questions first cannot supply. The simulation is also the most expensive round - two interviewers, one of them playing the customer, and a written summary the candidate sends afterwards - so if you need a cheaper filter in front of it, use a screen that asks nothing about discovery. In this loop its questions on being wrong and on what is left behind go in the interviewer's notes, because production judgment and ownership are each read twice elsewhere; its question about the team whose work the pilot would automate, its readback and its written summary are scored.

Round four is the AI Engineer loop's round on untrusted input and tool permissions. The assistant in it belongs to the interviewer's company rather than a customer's, and the questions need no adapting: what an agent may do on someone's behalf, and what happens on the morning it does too much, is the security review an FDE has to pass at every customer. Two things read differently here. Its pointer to a caching round means, in this loop, round three's index, where a voided work order kept answering. And its third section is an incident the candidate's own code allowed, which is where this loop reads ownership the second time.

Coding is a nice-to-have, read once, in round three - the same trade the AI Engineer loop makes. This loop will pass a candidate who writes plain code that survives a customer's estate over one who writes elegant code and has never had to ask for access. If your FDEs spend most of the week in your own codebase, add a general coding round and make Coding a must-have.