Overview / Technical
Four proprietary on-device models build a clinical draft from captured evidence. The veterinarian reviews and approves the record before sharing.
Separate evidence.
One case record.
Each recorded item retains its source. Spoken facts link to their sentence, speaker, time and interval. Monitor readings, initial selections and recording-clock values retain their own evidence or markings.
Procedure context guides what to listen for. It cannot supply a patient fact. Information that is missing remains Not documented.
Vetucate’s on-device models are trained on and guided with veterinary terminology and College of Veterinarians of Ontario record-keeping language.
Turns what the team says into a timed transcript. Veterinary vocabulary guides transcription. Missed speech is re-read, and unresolved passages keep their timing and an unclear marker. If punctuation repair fails, the original transcript is used.
Labels who is speaking. A final pass across the whole recording reconciles provisional speaker labels. Labels remain anonymous unless identity is established through the team introduction or selection workflow.
Reads heart rate, SpO₂, blood pressure and temperature from a visible monitor. Consecutive readings must agree within a clinical tolerance. Readings retain the source image region and the displayed temperature unit. Monitor and spoken values remain separate.
Writes the record from the transcript, guided by veterinary terminology, College of Veterinarians of Ontario record-keeping language and the selected procedure. Entries are restricted to the record structure and retain supporting evidence. It audits the case, checks professional wording and traces the narrative to recorded facts.
Each thirty-second interval is captured, blurred, fingerprinted and sealed. Extraction uses the preceding interval’s ending to retain sentences split across a boundary. Facts accumulate into a case record.
Restatements are merged. Conflicting values retain their evidence. The final audit may settle a conflict only when evidence settles it, with the reason recorded.
The record engine can fill only the fields in the fixed clinical record structure. Spoken entries require supporting quotations. Naming values are checked against captured words.
The narrative is written from record entries. Each sentence is checked against the facts it claims; unsupported sentences are removed and their removal logged. Missing facts are not supplied from procedure knowledge.
Completed intervals carry independent completeness information and a chained fingerprint. At Stop, the case state is persisted before final composition. Interrupted finalization can be rebuilt from that state.
527 procedures are catalogued across 14 species groups. 478 have researched profiles and owner-language explanations. Context remains separate from patient evidence.
Most wording checks run during the case. The clinical record uses professional wording while the timeline retains the spoken original.
Review the pipeline, measured performance and validation limits.
