Offtrackometer puts a wandering needle on your wandering meeting

Your meeting can now acquire a needle that points out how far it has wandered. Somebody was eventually going to build the conversational equivalent of the passenger watching your speedometer.

Max Harrison built Offtrackometer around an idea he’d had for years: a tool that notices when a meeting slips off topic. His November 2 write-up describes the decisions behind a working demo, including speech transcription, the amount of context to send and how to keep frequent model calls affordable. The page offers five free minutes without signing up, then £5 for 500 minutes.

The speech recognition runs locally. Harrison uses Chrome’s on-device recognition when available, with Moonshine Medium Streaming as the fallback. Moonshine downloads into the browser and runs through WebAssembly. That avoids sending the microphone stream to a transcription service, although Harrison says slower devices can struggle to keep up with real-time speech.

The words then take a different route. Offtrackometer bundles recent speech with the meeting goal and a little earlier discussion, and sends that context to Clef-flash on Cloudflare. The model returns a value between zero and one. The app smooths successive results and turns them into the percentage displayed by the dial. If you’re considering this for a real call, the distinction matters: local transcription still produces text that leaves the browser.

Sensitive mode examines about ten seconds of recent speech and waits at least three seconds after a response before checking again. Balanced mode looks at roughly thirty seconds and waits at least ten seconds. Those are two ways of deciding how quickly the passenger should interrupt you. A very short window catches the sudden tangent sooner, while giving the system less conversation with which to understand it.

Harrison chose a decision model because its constrained answer could go directly into the interface. He didn’t need a paragraph explaining why the meeting had become a discussion about something else. He needed a number the meter could consume, at a price and latency that made repeat checks practical. His write-up is the source for these observations, rather than an independent comparison of model quality.

Harrison writes that the definition of going off track is “not super clear.” A discussion can appear unrelated until somebody explains the missing dependency. Earlier context helps interpret references, but a small transcript window still has to distinguish a necessary detour from an enthusiastic account of someone’s weekend.

Harrison says the deployed demo doesn’t collect meeting transcripts. He also says it hasn’t had much testing and names evaluations against real meetings as a next step. That makes the toy useful as a prompt to talk about a call’s direction, with everyone’s consent. Turning its percentage into a performance score would ask the experiment to answer a question its maker hasn’t tested.

A future version might run a small decision model entirely in the browser; Harrison raises that possibility himself. First, somebody still has to establish what a good needle reading looks like when the apparent tangent is the reason the meeting succeeds.