The challenge
A warehouse walkthrough or a client workshop produces hours of spoken detail, and most of it is lost by the time anyone writes it up. Transcription services fix that, but they charge by the minute or by the month, and they want the audio uploaded to their servers.
Meanwhile most modern laptops carry a graphics card that sits idle all day and is very good at exactly this kind of work.
What we built
- One step. Plug the clip-on mic into the laptop. New recordings are copied, checked, transcribed and written up in the background.
- Transcription on the graphics card. A full-size speech model runs on the laptop's own GPU, so an hour of audio costs nothing and never leaves the machine.
- Speaker labels, also local. Each line is tagged by speaker, and names can be assigned afterwards.
- Tuned for real sites. Quiet voices are boosted, silence is skipped, and mixed English and Spanish is handled without switching settings.
- A report from every recording: summary, observations, decisions and actions. Only the transcript text goes to the AI service for this step, and it can be switched off for recordings that must stay on the laptop.
- A hard spending cap. The tool estimates the cost before every report and stops at the monthly limit. Over the limit, the transcript is still produced.
- Secure by design: keys held in the Windows credential store, copies verified before they count, no open network ports and no admin rights needed.
- A simple app window to browse recordings, tag them by client and site, and retry or re-run anything.
Results
Every site visit now ends with a searchable transcript and a structured report, without a subscription and without the audio leaving the laptop. The same approach works for any team sitting on capable hardware and paying per use for AI.
