WhatsApp AI support agent
A virtual persona that understands text and voice, answers with the clinic's real data and cuts manual support volume.
- Year
- 2025
- Status
- In production
- Context
- Multi-site pediatric clinic · Client under NDA
The challenge
The clinic received a high volume of WhatsApp messages across several locations. Repeated questions (insurance, hours, addresses, exam preparation) consumed the front desk, and after-hours messages simply went unanswered.
The solution
A conversational agent with its own persona, grounded in the clinic's real data: locations, specialties, accepted insurance plans and scheduling rules.
It handles both text and voice messages — audio is transcribed with Whisper before entering the flow, because a large share of patients send voice notes instead of typing.
Architecture
Evolution API bridges WhatsApp and pushes events by webhook into n8n, which orchestrates the flow.
Conversation history and context live in PostgreSQL, enabling continuity between messages and later auditing.
Inference runs on Groq (LLaMA 3.3 70B) for cost and latency; Whisper handles audio transcription.
Everything runs on a Linux VPS behind Traefik with automatic SSL, isolated in containers.
Result
24/7 first-line support with triage before reaching a human. Patient data stays out of the prompt, and demo screens have sensitive information removed.
What I learned
A good agent is a scoped agent: the more explicit its limits, the less hallucination and the less human rework.
Audio transcription isn't a detail — it's what makes real users adopt the tool.