Production Deployment · Home-Health Documentation

Clinicians talk
normally. The
platform writes the note.

ACE turns 90 minutes of charting into 5 minutes of review. Field clinicians record audio during a visit; the platform transcribes it, extracts the structured medical fields with Claude, and writes the result back into the Axxess EHR.

Industry
Home Health
Stack
Next.js · Expo · FastAPI · Claude
Status
Production · Live
EHR
Axxess (cookie-auth)
01 / Challenge

Clinicians spend more
time charting than
delivering care.

Home-health visit documentation is a regulatory bottleneck. PTs, OTs, and Nurses spend their evenings catching up on charting — vitals, M1850/M1860 functional scores, structured therapy data — fed by hand into Axxess, the EHR that runs the back end of the industry.

01
Charting eats the evening. A typical home-health visit produces 30–60 minutes of post-visit paperwork — fields buried across M1850/M1860 functional scores, vitals, therapeutic exercises, and a dozen Axxess form sections. Clinicians close their laptops at 9pm.
02
Generic dictation tools don't speak Axxess. Off-the-shelf voice-to-text produces a wall of text. Axxess needs structured fields — TherEx repetitions, pain scale, transfer assistance levels. The mapping is the work, and no consumer tool does it.
03
Axxess has no public API. No OAuth, no webhook, no clean integration path. Anyone writing back to it has to maintain a cookie-authenticated session against the legacy web forms — and refresh it before it expires.
04
Generic AI can't be trusted with PHI. Clinical documentation has to be auditable, profession-specific, and accurate enough that a clinician will sign their name to it. A general-purpose chatbot doesn't meet that bar.
02 / Solution

Record. Extract.
Write back.
Review.

A field-deployed mobile app plus a server-side AI pipeline that records, transcribes, extracts structured medical fields, and writes them into the EHR's actual form sections — not a transcript file, not a notes blob.

01
Record from the field
Expo SDK 53 mobile app for clinicians. TUS resumable upload of audio straight from the visit, with offline tolerance for the part of the day spent in the driveway between houses.
02
Profession-specific extraction
Anthropic Claude runs against profession-specific Pydantic schemas — PTVisitResponse, OTVisitResponse, NursingVisitResponse. What comes out is exactly what Axxess needs to see, not a transcript dump.
03
Writes back to the EHR
Per-section writers map structured output into 11+ Axxess form sections — Functional Mobility, TherEx, Pain, Start of Care, System Overview, Neuromuscular Reeducation. The cookie-authenticated session is refreshed on a cron and on demand.
04
Care plan goals from M-scores
M1850 and M1860 functional scores feed an automatic care-plan goal generator. The boilerplate that ate 20 minutes per evaluation is now zero.
05
Review surface, not approval theater
Next.js 15 dashboard with WaveSurfer audio playback (1×–2× speed). Clinician reviews, edits, signs — the workflow they already know, just without the typing.
06
Audit-ready by default
X-API-Key auth between services, a job tracker per visit, recording_url treated as sacred (preserve_recording_url=True everywhere). PHI handling considered from the architecture down.
03 / Before · After

What changes
for the clinician.

Before
Open laptop in the driveway. Type for an hour.
  • Visit ends, clinician opens laptop in the driveway.
  • 30–60 minutes typing into Axxess form sections.
  • Manual transcription of vitals, pain scores, exercises performed.
  • Care plan goals re-derived from scratch each visit.
  • Documentation backlog bleeds into evenings and weekends.
  • "Did I capture the right field?" — compliance anxiety as the default state.
After
Talk through the visit. Review. Sign.
  • Tap record. Deliver care. Tap stop.
  • Audio uploads via TUS while the clinician walks back to the truck.
  • Structured note ready in the dashboard within minutes.
  • M1850/M1860 → care plan goals, auto-generated.
  • Click approve — note posts to Axxess in the right form sections.
  • Compliance built into the schema, not chased after the fact.
04 / Built

By the
numbers.

7
tRPC routers
auth · patient · visit · careplan · dashboard · axxess · mace
11+
Axxess form sections
Functional Mobility, TherEx, Pain, Start of Care, …
3
Profession schemas
PT · OT · Nursing — Pydantic, validated end to end
2
LLM providers
Anthropic Claude (primary) · AWS Bedrock (fallback)
~5min
Charting → review
Down from a typical 30–60 minutes of post-visit typing
100%
Audit-trail coverage
recording_url · transcript · structured payload — every step persisted
05 / Architecture

Audio in.
Note out.

The pipeline is purpose-built for clinical extraction: audio in, profession-specific structured fields out, posted into the EHR through the only path it actually supports. Every layer falls under MACE — KyberCore's invisible operating system.

Surface
Expo Mobile · Next.js Web · Admin DashboardClinicians record on mobile, review on web; ops staff inspect raw visits in the admin UI
▼
Product API
tRPC v11 · better-auth · DrizzleAuth, patients, visits, care plans, and the Axxess client
▼
▼
Data
Postgres system of record · Axxess EHR (legacy spine)Structured fields posted into Axxess form sections via cookie-auth
I haven't typed a visit note in a month. I record what I do, the platform writes it up the way Axxess wants it, I read it, I sign it. The drive home is the drive home again.
PT, ACE Home Health · Production user · 2026
06 / Outcomes

What it delivers,
in practice.

01
Charting time becomes review time.Clinicians spend ~5 minutes reviewing and approving instead of an hour typing. The day ends when the visits do.
Clinician time
02
Profession-specific accuracy.PT, OT, and Nursing each get their own Pydantic schema. Extracted fields match the form, not a generic transcript.
Quality
03
Axxess write-back without an Axxess API.Cookie-authenticated form posting against 11+ sections, refreshed on a cron and on demand. Integration that the vendor never offered.
Integration
04
Care-plan goals from functional scores.M1850 and M1860 → generated goal text, ready for review. The repetitive narrative work disappears.
Workflow
05
Audit-ready, end to end.recording_url preserved on every update, structured payload persisted, every step logged. Compliance posture built into the data model.
Compliance
07 / Why it matters

Healthcare hasn't seen the AI productivity gains other industries are getting because the regulatory surface is real and the existing tools weren't built for it.

Generic AI dictation tools produce text. The work in clinical documentation isn't producing text — it's producing structured fields that map into a specific form, in a specific EHR, for a specific profession.

ACE is what AI-assisted documentation looks like when you take that surface seriously. Profession-specific schemas. EHR write-back through the only path the vendor actually supports. An audit trail built into the data model rather than bolted on after.

The same pattern that makes ACE work — schema-aware extraction, connector resilience, structured output validated end-to-end — is the foundation for MACE, the operating system underneath every KyberCore product going forward.

Run a home-health
operation?

30-minute demo with a real visit recording, transcribed, extracted, and written into a sandbox Axxess. No deck — just the product.