"Knot.ai" — Adaptive AI Interview Panel

Interview practice today means reading questions off a screen. Real interviews are spoken, they get harder when you do well, they pin you down when you are vague, and they are run by several people who each want something different and compare notes afterwards.

The obvious build — hand the whole interview to an LLM and ask it nicely — failed in testing. Against an LLM-driven candidate, the interviewer model tunnel-visioned on one topic and escalated difficulty freely, ignoring “adjust by one level” and “cover three areas” the moment they competed with its own instincts.

I split the job. The model does what only a model can: understand an answer, judge it, and phrase the next question in character. Everything countable became a state machine in an OpenAI-compatible proxy that Agora’s voice engine calls on every turn — difficulty is a streak-based ladder clamped to 1–5, topic coverage rotates on a counter, and the right/partial/wrong verdict is emitted as an out-of-band tag that the proxy speaks itself, so what the candidate hears and what the report tallies can never disagree.

A panel of five personas (technical, hiring manager, stakeholder, product, behavioural) hands off between rounds through a short digest — covered topics, grounded bullets, end difficulty — which is the only thing that crosses into the next interviewer’s prompt. Afterwards, the report can only say what the candidate actually said: every strength and weakness carries a quote, and a verification pass drops any claim whose quote cannot be found in the transcript.