An interviewer-side signal layer for live mock interviews, for about 1.5¢ per interview-hour. From a live transcript, it produces a few judgments, built on TypeSafe's Jev, which returns labels and probabilities (instead of prose). The replay below shows both what was shown and what was dropped — and what it cost, ticking up as it runs.
0:00
Transcript
What the interviewer sees
—
Nothing yet.
Judgments made
0
Shown
0
Suppressed
0%
Median latency
0ms
Cost so far
0.00¢
How it works
- Typed, not prose: every few seconds a rolling window of transcript becomes a phase choice plus bounded scores for answered-question, depth, clarity, and rambling risk. Values, not paragraphs, so thresholds and rate limits are just code.
- Suppression is the product: 111 of 119 judgments in this session were discarded — below threshold, unchanged, low-confidence, or too soon after the last card. A correct signal can still be bad product behavior.
- Some questions never reach the model: preconditions in code — no open question, too little candidate speech, a turn too short to ramble — blocked 81 of 200 questions in this session, 40% of them, before a call was made. Duration and counting stay in code too, because Jev 1.13 does not count reliably; TypeSafe's own guidance is that extraction is a judgment and arithmetic is not.
- Cost: a quarter of a cent for the whole 10-minute interview — 39 calls, 60k input tokens, at Jev's $0.042 per million input tokens with output free. About 1.5¢ per interview-hour, at a median 119 ms per call, which is what makes a per-tick cadence viable at all.