Patient Message Triage API

An AI routing service: a message with a refill and a new symptom is split and sent to the nurse and clinician queues. It suggests a route, never a diagnosis.

AI that reads patient messages and sends each request to the right team, in under 2 seconds.

Patient messages often mixed a refill request with a new symptom. One label sent the whole thread to one desk, and urgent symptoms waited behind billing.

One message often held a refill and a new symptom. A single label sent the whole thread to the wrong desk, and urgent swelling waited behind billing.

• •Split mixed messages into separate intents.
• •Route each intent to a queue the clinic actually staffs.
• •Return a route, never a diagnosis or treatment.

I built the Node.js service, the prompts for Claude and ChatGPT, the closed label set and RBAC map, and the Redis routing cache.

A Node service calls Claude or ChatGPT with a fixed label set, stores the route in Redis for the shift, and refuses labels outside the clinic’s RBAC map.

1. 01

   REST endpoint

   Accepts a patient message and returns one route per intent.

2. 02

   LLM classification

   Claude or ChatGPT splits intents and labels each from a fixed set.

3. 03

   RBAC label map

   Filters out any label the clinic has not mapped to a staffed queue.

4. 04

   Redis

   Stores routes for the shift so queues read quickly.

• •Routing in under 2 seconds.
• •Closed label set enforced after the model.
• •Automatic first-pass inbox sort.

const labels = ['refill', 'symptom', 'billing', 'scheduling'] as const;

export async function triage(message: string) {
  const raw = await llm.json({
    system: 'Split intents. Use only the allowed labels. No diagnosis.',
    input: message,
    schema: { intents: labels },
  });
  return raw.intents.filter((item) => labels.includes(item.label));
}

1. 01

   Mixed intents had to split instead of picking the first sentence.

   The prompt asks for all intents, and the schema expects a list, so a message with a refill and a symptom produces two routes.

2. 02

   The model could not invent a queue the clinic did not staff.

   Model output is filtered against the clinic’s RBAC map; unmapped labels are refused.

3. 03

   A triage suggestion is not a diagnosis, so the API returns a route, not a treatment.

   The API returns a route only and the prompt forbids diagnosis, keeping triage separate from treatment.

Security

RBAC controls which queues exist per clinic, and only routes are stored, not clinical conclusions.

Speed & growth

Each message routes in under 2 seconds, with shift routes cached in Redis.

Message fixtures with mixed intents, out-of-set labels, and requests for advice, each with expected routes.