What we learned testing outside software engineering

Almost every public demo of voice interview AI features a software engineer. It's an easy audience to impress — the vocabulary is consistent, the resumes follow a familiar shape, and the questions ("tell me about a system you designed") tolerate some vagueness in the model's follow-ups without anyone noticing. It's also the least useful test of whether the system actually works, because it's the case the underlying models were most exposed to during training.

We ran the Voice and Evaluator engines against candidates deliberately chosen to be nothing like that: an aerospace design engineer, a farm operations supervisor, an ICU registered nurse, and an office administrator. None of them work in an industry a general-purpose language model has seen described in the same density as software engineering. That was the point.

Aerospace: anchoring to a resume that isn't generic

The aerospace interview ran the full twenty questions, every one of them anchored to specifics from the candidate's actual resume — a named project, a real employer, a specific system. A resume-anchored interview is only as good as the model's ability to ask a real follow-up about a real detail, not a generic version of the question that would work for any resume. That held up.

Agriculture: the reconnect test

Voice sessions drop connections. A farm operations supervisor's interview reconnected mid-session, and the engine picked up exactly where it left off — zero question repeats across the reconnect. That's the state-machine discipline from the Voice Engine doing its job: the model doesn't get to "forget" where it was just because the socket did.

Healthcare: the pause that matters

An ICU registered nurse's closing sequence held a genuine sixteen-second wait for a real reply before moving on. Sixteen seconds is a long time for a system that's optimized, by default, to keep a conversation moving — and it's exactly how long a real hiring manager would actually wait before assuming someone was done talking.

Administrative: the boring case that still has to work

An office administrator's interview is the least dramatic of the four, and that's exactly why it's a good test — there's no interesting technical depth to hide behind, just accurate anchoring to a real employer and real tool history, question after question, without the model reaching for a more "interesting" generic answer instead.

The industries above already validate the interview engine works generally — the honest next step is extending the same rigor to matching and screening ahead of the interview, not just during it.

None of this is a claim that the system is finished. It's a claim that it's been tested somewhere other than the case that was easiest to get right.