- Client
- Cadence Health
- Industry
- Healthtech
- Year
- 2024
- Duration
- 18 weeks
- Team
- 4 engineers, 2 designers, 1 clinical researcher
- Services
- Product design · AI engineering · Web platforms
Clinicians at Cadence spent eleven hours a week charting. We rebuilt the clinical workspace around ambient capture and a hard human-review gate, and handed most of that time back.
The product
Rebuilt interface, live figures. Everything below is the shipped surface, not a concept.
Note drafted
92%
Awaiting review
4 fields
Time in note
6:24
Ambient transcript
- 04:12Presenting with a three-day cough, no fever.
- 06:38Chest clear on auscultation, sats 98 on room air.
- 09:38Impression is viral, no antibiotics indicated.
Drafted fields
Nothing reaches the record until accepted
Rm 4B · 12 min elapsed
Encounter 8841
Nothing is written until accepted
Bedside tablet — review queue
The charting flow was fourteen screens deep and had plainly been drawn around billing codes rather than around a person holding a tablet at 2am. Clinicians were logging eleven hours a week in documentation and finishing 38% of their notes after the shift ended. Every change had to clear HIPAA review first, and the Epic instance underneath was not ours to modify — not difficult to modify, not ours.
We shadowed nineteen clinicians across three sites before drawing a screen, and that is where the fourteen screens turned out to be a symptom: people were re-entering the same three facts because no screen trusted the one before it. The workspace collapsed into a single scrolling encounter view with structured templates per visit type, and an ambient model drafts the narrative sections while the clinician talks. Every drafted field lands in a review state with its source utterance attached — nothing reaches the record until a human accepts it, and we held that line in month one when it was costing us adoption. Inference runs on GPU nodes inside the tenant VPC behind PrivateLink, so PHI never crosses the boundary.
Time per encounter note fell from 13.4 minutes to 6.4, about 5.8 hours a week back per clinician. After-shift charting dropped from 38% of notes to 9%. The workspace reached 94% weekly active adoption across 32 clinics inside two quarters and cleared a third-party HIPAA assessment with no remediation items. The complaint we get now is about template coverage for specialties we have not built yet, which is the complaint we were hoping for.
Solid lines carry request traffic. Dashed lines are asynchronous — replication, shadow reads and audit trails that must never sit on the critical path.
export function commit(field: DraftField, actor: Clinician): ChartField { if (field.state !== "accepted") { throw new UnreviewedField(field.path) } return { value: field.value, provenance: { utteranceId: field.source.utteranceId, model: field.source.model, confidence: field.source.confidence, }, acceptedBy: actor.npi, acceptedAt: field.acceptedAt, }}The review gate is one function, and it throws. Nothing reaches the record without a clinician's accept, and every accepted field keeps the utterance it came from.
Less time per encounter note, 13.4 minutes down to 6.4
Charting hours returned per clinician each week
Weekly active adoption across 32 clinics, two quarters after launch
Figures reported by Cadence Health after launch.
- 01Single-view encounter workspace replacing a fourteen-screen flow
- 02Draft-and-review model with source-utterance provenance on every field
- 03In-VPC inference over PrivateLink so PHI never leaves the tenant boundary
- 04FHIR write-back against an unmodified Epic instance
- 05Offline-tolerant tablet client for low-connectivity floors
- Next.js
- TypeScript
- Python
- FHIR / HL7
- PostgreSQL
- AWS PrivateLink
“I stopped taking charting home. That is the whole review — I get my evenings back, and the notes are better than the ones I used to write at midnight.”
