Notes from shipping a voice mock-interview tool

Field notes, kept because the same problems kept recurring.

I have been building MakeInterview, a voice-first mock interview practice tool for candidates who prepare for English-language technical interviews in a second language.

Reading answers does not transfer to speaking

The thing that surprised me most is how poorly silent preparation transfers to a live spoken answer. You can know the shape of a good answer and still produce a shapeless one at speed. So the tool runs an actual spoken interview loop and then returns a scorecard on fluency, structure and content, rather than a page of prompts.

What actually moved the needle

Three things mattered more than expected:

Latency. Past roughly 800 ms between the candidate finishing and the interviewer replying, the illusion of a conversation collapses and people start performing for a machine again.

Transcription drift. A single mis-heard technical term turns the follow-up question into nonsense, and the candidate usually blames their own English rather than the transcript.

The replay. Hearing your own answer next to a model answer improves the next attempt far more than reading a score.