A sick person is a poor historian — they forget, they minimise, they rehearse. The cabin reads face, voice and symptom before a single question is asked, so the patient arrives already understood rather than interrogated.
Every signal is traceable to the model that produced it. The clinician decides; the machine only ever hands them evidence.
Most telemedicine is a funnel — book, call, leave. CabinOS closes: what the cabin learns today becomes context for the next visit.
The cabin greets the patient and takes consent before anything is recorded.
A single frame reads visible signs of distress. No video is kept.
Symptoms in the patient's own words, plus tone, stress and speech rate.
A weighted score places them in the queue by clinical need, not arrival time.
The doctor joins already holding the report — and a copilot that listens.
A drafted SOAP note the clinician edits and signs. Never auto-finalised.
Then it remembers. The completed visit is embedded and stored, so the next time this patient walks in, the doctor sees what changed since last time — without asking them to remember it.
Four signals, each from a real model, each shown to the clinician with its provenance attached.
Seven-class expression read from one captured frame, returned with a confidence score and a clinical note.
Vision transformer
Valence, arousal, vocal tremor, speech rate and pain likelihood — the things a patient rarely reports accurately.
Speech emotion model + DSP
Symptoms, medications and conditions lifted from the transcript as structured terms a doctor can scan.
Medical NER
Past visits retrieved by meaning, so “heart attack” finds “myocardial infarction”. Empty on a first visit — never invented.
Medical embeddings + vector search
Seen from where you stand in it.
Medical software earns trust by what it refuses to do. These are constraints in the system, not promises in a policy.