Artificial IntelligenceIn progress
AI Clinic Assistant
Trợ lý Phòng Khám AI
AI receptionist that answers calls, books appointments and sends reminders — integrated with Twilio, doctor schedules, in-house EMR.
PythonTwilioOpenAI RealtimeFastAPIPostgreSQLWebSockets
Problem
A private clinic missed 30% of peak-hour calls because the receptionist couldn't answer + check the schedule + book appointments fast enough. Patients bounced to competitors. Lost revenue ~12% monthly.
Architecture
Twilio Stream → WebSocket → OpenAI Realtime API (voice-to-voice) → function-calling to EMR (check slot, book, send SMS confirm). Guardrail prompt: only book — never prescribe or diagnose.
Stack & rationale
- OpenAI Realtime: ~600 ms end-to-end voice latency, far more natural than TTS+STT chains.
- Python WebSocket: in-process, no queue hops → low latency.
- Twilio Vietnam: local number, ~~$0.01/min.
Results (3-month pilot)
| Metric | Before | After |
|---|---|---|
| Peak-hour answer rate | 70% | 100% |
| No-show (with reminders) | 22% | 17% |
| Daily successful bookings | 14 | 24 |
Lessons
"No prescription" guardrail must live in both system prompt and tool descriptions. For real production, add an "escalate to human" tier for edge cases.