coconutlabs

gallery unit · applied ai

CareScribe

A clinician talks for four minutes. The chart needs five headings, and every one of them has to be supported by something that was actually said.

scenario sandbox

You are the clinician at the end of a visit. The transcript is what was said. The note is what has to go in the chart, and it has to be right.

pick a transcript

READY

Pick a transcript or write your own, then run it.

382 chars · cap 4000

Synthetic transcripts, written for this page. Do not paste real patient data: it leaves your browser to reach the model. Nothing is stored at either end.

what you are looking at

  1. 01 · transcript in

    A synthetic visit transcript, either one of the three presets or whatever you type.

  2. 02 · model runs

    Your browser posts it to a Worker. The Worker calls a model on Cloudflare's GPUs and waits. The note comes back whole.

  3. 03 · note out

    Five headings. Anything the transcript does not support is marked not documented rather than filled in.

Go deeper

The instruction the model is given

The prompt is short and most of it is prohibition. It fixes the five headings, then spends its remaining words telling the model what not to do:

Write only what the transcript supports. If the transcript does not
cover a heading, write 'not documented' under it. Never invent a vital
sign, a dose, a lab value, or a diagnosis.

Run the thin transcript preset to see whether it holds. That transcript supports almost nothing, and a model that wants to be helpful will happily produce a full note from it. The interesting output is the one that mostly says not documented.

How this runs for nothing

Workers AI is included in the Cloudflare free plan with an allocation of 10,000 neurons per day. The Worker in front of it is on the free plan too, which caps a Worker at 10 ms of its own CPU per request. That sounds fatal for a model call and is not: CPU time counts only the Worker executing its own code, and time spent waiting on a binding does not count. The Worker validates, prompts, waits, and shapes the reply, which costs a couple of milliseconds. The model does the work on hardware I do not pay for.

When the daily allocation runs out the endpoint returns 503 and says so, rather than failing in a way that looks like a bug. It resets at 00:00 UTC.

What this demo does not show
  • No real patient data, ever. The transcripts are synthetic and written for this page. The repository behind it is a curation pipeline over public CC BY 4.0 corpora with staged de-identification, and carries no PHI.
  • No evaluation. This page shows one model answering one transcript. It does not show accuracy against a reference, agreement between clinicians, or a hallucination rate. Those numbers do not exist yet and the page does not imply them.
  • No fine-tuning. This is a general instruction-following model with a careful prompt. Any claim that it was adapted to clinical text would be false.
  • Not a medical device and not clinical advice.
Source All units