The problem

Every transcription service asks you to upload the recording. For a therapist, a lawyer taking a deposition, or a doctor dictating notes, that upload is the entire problem. It turns a private conversation into someone else’s data, sitting on someone else’s disk, under someone else’s retention policy.

What Lucy does

Captures the session, transcribes it, works out who said what, and produces five outputs: a summary, action items, decisions, key topics, and a Q&A. Then it does all of that without a network connection.

The craft

Speech recognition runs through WhisperKit on the device’s neural engine. Speaker separation uses FluidAudio. The summaries come from open-weight language models, Mistral 7B and Phi-3-mini, quantised to four bits and run on the device’s own silicon.

That combination is the point. There is no API key, no inference endpoint, and no request leaving the device, which means the privacy claim is a property of the architecture rather than a promise in a policy document.