Design a clinical documentation assistant for doctors in a hospital (ambient scribe), considering privacy and safety.

Pipeline

1. Capture: mobile or room device; explicit patient consent; start/stop control by the clinician.
2. Transcription: medical-domain STT; handling of multilingual consultations (e.g. Hindi/English mixes); speaker labels (doctor vs patient vs attendant).
3. Note generation: an LLM with the hospital's note template and specialty-specific prompts; it uses only what was said in the encounter plus permitted EHR context (medications, allergies); it marks uncertain items.
4. Review UI: side-by-side transcript and note; click a sentence to see the source utterance; quick edits; sign-off.
5. EHR integration: write the signed note via standard interfaces (e.g. HL7/FHIR); no automatic orders or prescriptions.

Safety and compliance

• Hallucinations and omissions are the main risks. Measure them with clinician-graded evals (per-section accuracy, critical omissions, fabricated findings).
• Scope limits: documentation only; no autonomous diagnosis or treatment suggestions in v1.
• Privacy: PHI protections under applicable law (e.g. DPDP Act in India, HIPAA in the US), encryption, access control, audit, retention limits for audio (often delete after note signing).
• Clinical governance: a clinical safety officer or committee approval, an incident reporting process, monitoring of edit rates by specialty.

Metrics: documentation time per encounter, after-hours charting time, note quality scores, clinician satisfaction and adoption.

Related

• Previous: Q39. A manufacturer wants an AI assistant for procurement: comparing supplier quotes, checking contracts, and drafting purchase orders in their ERP. Design it.
• Next: Q41. Case interview: "Our customer support costs are too high. How can AI help?" Walk through your full answer.