Clinical Notes AI API

An AI service that reads a clinical note and fills in the fields the clinic already uses. Every value must come from the note itself, so no invented doses are ever stored.

AI that turns a doctor’s free-text note into structured problems, medications and follow-ups.

Clinicians wrote a note, then typed the same history again into the discrete fields that billing and follow-up screens required.

Clinicians typed the same history twice: once in the note, again into discrete fields the billing and follow-up screens required.

• •Turn free-text notes into structured problems, medications and follow-ups.
• •Store only fields the clinic already uses, and nothing the note does not support.
• •Keep patient text out of logs.

I designed and built the Python extraction service, the closed schema and validator, the LLM prompts for Claude and Gemini, and the logging policy.

A Python service sends the note to Claude or Gemini with a closed schema, then rejects any key that is not problems, medications, or follow-up.

1. 01

   REST endpoint

   Accepts a note and returns structured JSON in a fixed schema.

2. 02

   LLM call

   Claude or Gemini with a system prompt that asks for the schema only and forbids invented fields.

3. 03

   Python validator

   Parses the response, drops unknown keys, and rejects any value without a source span in the note.

4. 04

   Chart write

   Only validated problems, medications and follow-ups reach the chart.

• •Closed-schema extraction with zero invented doses stored.
• •One API path for every note.
• •Patient text excluded from application traces.

async function extractClinicalNote(note: string) {
  const completion = await llm.complete({
    model: 'claude',
    system: 'Return only the clinical JSON schema. Do not invent fields.',
    input: note,
  });
  return ClinicalNoteSchema.parse(JSON.parse(completion.text));
}

1. 01

   The model would invent a dose that was not in the note. The parser drops any value without a source span.

   The parser drops any value that has no source span in the note, so an invented dose never reaches the chart.

2. 02

   Notes mixed English shorthand and abbreviations. The schema had to accept both without guessing a diagnosis.

   The schema accepts shorthand and abbreviations as text, without inferring a diagnosis the note does not state.

3. 03

   Patient text could not be logged in application traces.

   Logging records request metadata and validation outcomes only, never the note content.

Security

Notes are processed in memory and never logged. The schema is closed, so the model cannot add fields that would carry extra patient data into storage.

Speed & growth

One synchronous extraction call per note, with validation in Python adding negligible overhead to the model call.

A fixture set of notes with expected outputs, including notes designed to tempt the model into inventing doses or diagnoses, runs against the validator.