Intelligent Interview Analyzer

A summary of what somebody said is not a finding. Ask a model to read an interview cold and it hands back a competent précis of the conversation. What you needed was what that hour tells you about the question you were asking, and the gap between those two is context.

So it takes the study with it, not just the transcripts. The interview plan, what was being tested, and the material that says what this round was for all go in, and the prompting is built around holding them while it reads. A transcript is measured against something rather than merely read, and what comes back points somewhere.

Then there is doing it fifty times

A whole round together is a second problem on top of the first, because it does not fit anywhere you could put it at once. Sent in one piece it is refused, quietly truncated, or so thin the synthesis is worth less than the notes.

So the core is built for the volume. The work is pooled instead of attempted in a single pass, and it is tested at thirty to fifty transcripts in a run, repeatedly, with no failed run. One command starts the server, runs its own health check and diagnostics, and opens the app.

A whole round of transcripts is measured against the study, the interview plan, what was being tested and what this round was for, then pooled across a core built for the volume into one synthesis. All of it on your machine: nothing is uploaded.

Your machine

Nothing is uploaded

1. A whole round

   It does not fit anywhere you could put it at once.

2. The study

   The interview plan, what was being tested, and the material that says what this round was for.

3. A core built for the volume

   The work is pooled instead of attempted in a single pass.

4. One synthesis

   Over the whole round, actionable rather than a précis.

Whose analysis it is

The team already paid for a research tool with AI in it, so cheaper was never the reason, and neither was smarter. A general tool imposes a general method. The interesting part of a research practice is the specific part: which cuts you take, what counts as a signal, how you decide two people said the same thing. Loading the study in is how you tell it.

And it stays on your machine

It runs against your own key and nothing is uploaded. Interview recordings are personal data, and the people in them did not agree to sit on somebody's server because it was convenient for whoever was analysing them.