Nora Health

Nora Health listens to a clinical consultation and produces the note a clinician would have written afterwards. Transcription, structuring, and drafting all run on the clinician device, so patient audio is never uploaded, never stored in a third party account, and never used to train a model elsewhere.

The product exists because the documentation burden in healthcare is measured in hours per clinician per day, and because the obvious solution, streaming consultations to a hosted model, is the one most clinical governance teams will not approve.

Clinicians spend a significant part of the working day writing notes rather than seeing patients. The available automation almost all depends on sending recorded speech to a cloud service, which puts identifiable patient data into someone else data centre and turns a clinical tool into a procurement and information governance problem.

The second problem is quieter and more serious. A language model asked to summarise a consultation will produce a fluent note whether or not it understood the conversation. A note that reads well but records a medication the patient never mentioned is more dangerous than no note at all, because a fluent error invites trust.

The system runs entirely on the device. Speech recognition is a quantised model small enough for consumer hardware, and note generation is constrained so every clause in the output has to trace back to something that was actually said.

Rather than asking a model for a note and hoping, the pipeline extracts candidate clinical facts first, grounds each one against a span of the transcript, and only then composes prose from the facts that survived. Anything the model asserts without a supporting span is dropped before a clinician ever sees it, and the interface shows which part of the conversation produced each line.