Accurate multilingual interpretation
Real-time interpretation that stays accurate on clinical terminology, disordered and pediatric speech, and rapid language-switching — where general-purpose translation breaks down.
Research
Clinical AI that works in any language — and knows when it might be wrong.
Allied-health therapy happens through conversation. When clinician and patient don't share a language, care suffers. We're building a calibrated, multilingual clinical-AI pipeline for PT, OT, speech and ABA practices — and measuring it like a research program, not a feature list.
What we are measuring
Our research aims
Three questions, each with a status that reflects where the evidence actually stands — not where we hope it will.
Calibrated clinical-safety confidence
A per-utterance confidence signal that reliably flags unreliable output, so a clinician catches an error before it reaches a patient or a claim.
Documentation & coding accuracy
Turning the encounter into structured notes and confidence-scored codes — surfacing likely problems before a claim is ever submitted.
Evidence
Evidence to date
Verified results publish here as validation completes — every figure dated, with its sample size and a methodology note. Until then, the live telemetry below fills in automatically as the pilot runs.
Live pilot telemetry
Auto-computed from de-identified activity streams — each figure appears on its own the moment its stream reaches five data points, and graduates to a verified result above once methodology review completes.
Method
Our approach
Measured the way a clinical study is measured: against real encounters, graded by clinicians, with the failure modes counted.
Real encounters
De-identified evaluation corpora of real allied-health encounters, graded by clinical experts.
Clinical error rate, not word error
Accuracy measured against general-purpose baselines by clinically significant error rate.
Calibration that predicts
Confidence validated so that a low-confidence flag genuinely predicts a real error.
Clinician in the loop
By design: the system assists, a human decides.
Responsible AI
Safety and privacy are design requirements
Sefton is built for a clinical setting, so trust is not an afterthought.
HIPAA-aligned by construction
Research metrics are computed from counts, hashed identifiers and categorical outcomes — never patient records.
Aggregate, minimum sample, named reviewer
Published figures are aggregate-only with a minimum sample size, verified by a named reviewer before they appear here.
Measurement only
Every AI suggestion is measurement-only with a human in the loop; nothing bills or acts on its own.
Programs
Programs & recognition
Labels reflect current application status, not endorsement. Updated as programs progress.
Work with us
Partner on research, run a pilot in your clinic, or reach us for media. Preprints and white papers forthcoming.