twenty80.studio
Book a call

Case study · Product

Built like infrastructure.
Reads like the doctor.

Every line traceable. Every note theirs.

Scribes are everywhere. Notes doctors trust are not. We built one into a healthcare product that knows the patient, writes like the doctor, and shows where every line came from.

Before

  • 36 minutes of EHR work per visit.
  • Another 1.4 hours at home, most nights.
  • Generic scribes helped, but notes still needed rewriting.

We audited it, then split it three ways.

  1. Integration

    Reads and writes the EHR

    One data model for every EHR, from Epic and Cerner to athenaHealth and AdvancedMD. Before the visit it reads history, medications and past notes; after sign-off it writes the note, problems, medications and orders back.

  2. Templates

    Writes like the doctor

    A template per specialty, tuned to each doctor’s own past notes. The structure is fixed by the template, not guessed by AI.

  3. AI

    Drafts, and shows its work

    It drafts the note from the conversation, with the history as context. Every line links to the moment it came from. Anything unsure is flagged, not guessed.

How it’s built

Six services. Each scales on its own load.

A visit is not one workload. Audio streams live, speech arrives in bursts, drafts pile up as visits end, and the EHR has its own speed limits. So each part runs as its own service on Kubernetes, behind one API gateway, linked by queues, and sized to its own curve.

  1. Visit capture

    Streams audio from the room.

    One stream per live visit

  2. Transcription

    Medical-grade speech model on GPUs.

    GPU pool, sized by minutes of audio

  3. Context

    History, medications, allergies, last visit.

    Fetched overnight for the day’s list, cached

  4. Drafting

    Writes the note into the doctor’s template.

    Queued per finished visit

  5. EHR write-back

    Files the note and structured fields.

    Paced to the EHR’s limits, safe retries

  6. Audit

    Logs every read, draft and signature.

    Append-only, always on

Quality

Quality is engineered, not hoped for.

Every line in a note has to hold up. So every line can show where it came from.

General versus medical-grade speech models, from a published benchmark.

  1. Medical-grade transcription

    A speech model trained on clinical language. Drug names and doses are checked against the patient’s medication list.

  2. Context before words

    History, medications, allergies, labs and the last visit are found by keyword and by meaning, re-ranked, and summarised before the patient walks in.

  3. Compliance in every prompt

    Nothing the doctor didn’t say. Hedging kept as spoken. Only the patient data the note needs.

  4. Release gates

    Each release is scored against notes doctors signed, per specialty. An omission or a made-up line blocks it.

What changed

95%
of drafted notes accepted by doctors
99%+
accuracy retrieving the right patient history
1 week
to connect a new EHR, down from 3 months
  • Live with 36 paying pilot clinics, used by more than 60% of their clinicians.
  • Reviewed and signed in one screen, inside the EHR. No new app.
  • HIPAA from day one: agreements with every vendor, no data kept by model providers, everything encrypted and logged.

Stop piloting.Start compounding.

It starts with the map: where AI pays in your company, what it costs, and the first build worth doing.

Or write to [email address]