Example: Patient Intake & Triage

​

What you’re building

​

Step 1: Create the project

​

Step 2: Describe what you want

“I need a patient intake form that patients can fill out without logging in. It should collect their name, date of birth, chief complaint, and whether they have any known allergies. When they submit, an AI should classify the urgency as low, medium, or high. If high, page the on-call clinician immediately. The clinician should see a dashboard with all recent intakes sorted by urgency.”

• Creates a patient entity with fields: name, date_of_birth, chief_complaint, known_allergies, risk_level
• Creates a Public app (intake form) at /intake
• Creates a Role-gated app (clinical triage board) at /triage with audience: clinical staff
• Creates an intake workflow with trigger: form submission
• Compiles a workflow with steps: receive payload → PHI detection → risk classifier (AI inference) → conditional branch → notification

​

Step 3: Review the contract

Trigger: user_interaction (form submission, app: patient-intake)

Steps:
  1. phi_scan        — scan payload for PHI; anonymize for inference
  2. risk_classify   — AI inference: classify urgency from chief_complaint
                       output: { risk_level: enum[low, medium, high] }
  3. entity_write    — write patient record with risk_level
  4. condition       — if risk_level == "high"
     ├─ true:  notify_oncall — Postmark integration → on_call_clinician_email
     └─ false: (no action)

PHI policy: strict

​

Step 4: Add the Postmark integration

“Use Postmark for the email. I’ll add the API key now.”

​

Step 5: Configure the setting

on_call_clinician_email = [email protected]

​

Step 6: Test the intake form

• Name: Jane Doe
• Date of birth: 1985-03-15
• Chief complaint: Severe chest pain radiating to left arm, shortness of breath
• Known allergies: Penicillin

​

Step 7: Inspect the trace

1. phi_scan — detected PHI in name, date_of_birth. Anonymized for inference.
2. risk_classify — AI classified risk_level: high (confidence: 0.97). Inference ran on anonymized payload; re-identified for record write.
3. entity_write — patient record created. Record ID: pat_01JA7QG2...
4. condition — risk_level == "high" → true
5. notify_oncall — sandbox: “would have sent email to [email protected] with subject: High-risk intake: chest pain.”

​

Step 8: Test the triage board

​

Step 9: Release

• Summary: “First release: patient intake form, clinical triage board, and high-risk notification workflow.”
• Diff: 2 new apps, 1 new workflow, 1 new entity (6 fields), 1 new integration, 1 new setting.
• No schema migrations needed (first release).

​

What happens in production

1. A real patient submits the intake form.
2. PHI detection scans the payload. PHI is anonymized before the AI inference call.
3. The risk classifier runs. Output: risk_level: high.
4. Patient record is written with risk_level: high. PHI is stored encrypted in the entity record.
5. Postmark sends the real alert to [email protected].
6. The clinician opens the triage board. The new intake appears at the top, risk badge: High.
7. The full trace is available in the audit log. PHI-redacted view for standard users; PHI-view role required for raw data.

​

Extending the example

• Add a follow-up workflow triggered when the clinician marks the intake as reviewed (data-change event on patient.status).
• Add a discharge summary workflow triggered on discharge — AI generates a summary and posts to the billing system.
• Add appointment scheduling — integration with a scheduling system, triggered after clinician review.
• Require 2 approvers for all releases — set in Settings → Release Policy (or as the org floor).