Max Harrison built Offtrackometer to flag when a meeting drifts away from its stated goal. His writeup describes the speech pipeline, context windows and update pacing. Despite the Jev headline, he says the deployed classifier is Cloudflare’s Clef-flash.
Keep the judgment focused on recent speech
- Harrison uses Chrome’s on-device speech recognition, with Moonshine Medium Streaming running locally through WebAssembly as a fallback.
- The hosted classifier receives the meeting goal, earlier discussion and recent speech. Local transcription therefore still leads to remote text processing.
- His prompt asks whether the latest discussion helps the goal. Earlier discussion should resolve references without outweighing the latest passage.
- Harrison describes sensitive mode as roughly ten seconds of speech, with at least three seconds between a response and the next request.
- Balanced mode uses roughly thirty seconds, with at least ten seconds after a response. The app smooths the returned value before moving its needle.
Test the meaning behind the meter
Harrison says integration was easier and felt more responsive than his earlier Gemini Flash attempt. He provides no matched timing or accuracy evaluation, and acknowledges that small snippets can make useful detours look irrelevant.
The displayed percentage comes from a typed Noul value, a probability-like answer to a true-or-false question. A valid typed value does not establish correct topic detection or calibration on meetings. Smoothing can steady the display while also changing when a detour becomes visible.
Caveat: the article was inspected; the app and its performance were not independently tested. Harrison proposes real-meeting evaluations and an entirely local Kev version as future work.