A clinical scribe where the audio never leaves the building.

The situation

Clinical documentation is the archetypal AI use case: a practitioner spends a significant part of every day converting a conversation they have already had into a structured record. The task is well defined, the input is speech, the output is a document with a known shape, and nobody enjoys doing it.

It is also the archetypal AI blocker. The input is a recording of a consultation, and in many practice settings the acceptable answer to "where does that audio go" is nowhere. That constraint kills a great many otherwise sound proposals — usually at a review meeting, months into a project, when the architecture can no longer accommodate it.

What we built

The product is a mobile-first web application, because the work happens standing up in an examination room rather than at a desk. A practitioner records the consultation, and the recording is sent to a speech-to-text service running on local hardware — not a cloud speech API. The audio does not leave the environment the practice controls.

The resulting transcript, which contains no voice biometrics and can be reviewed before it goes anywhere, is then used to generate a structured clinical note in the format the profession expects. That step uses a hosted language model, and the split is deliberate: the most sensitive artefact — the recording itself — never travels, while the less sensitive artefact goes to the best available model for the reasoning task.

Self-hosting is genuine rather than nominal. The application runs on hardware the practice owns, reachable over a private network tunnel with a real certificate on a real domain — necessary because browsers will not grant microphone access over an insecure connection, which is exactly the kind of practical detail that separates a demonstration from a product somebody can actually use.

Why it matters beyond veterinary practice

The interesting result is not the clinical note. It is that the architecture is transferable. Any workflow where speech is the natural input and the recording is the sensitive asset — legal intake, incident debriefs, field inspection, safeguarding interviews — has the same shape and the same blocker.

Splitting the pipeline so that the sensitive step runs locally and the reasoning step runs wherever it is best served turns a hard governance objection into an ordinary architecture conversation. That is usually worth more than any individual product built on top of it.

The general lesson

Data residency is treated as a constraint that rules use cases out. Far more often it is a constraint that determines where each step of the pipeline runs. Deciding that deliberately, early, is considerably cheaper than discovering it in a review after the architecture has set.