DermoAI

AI employee for clinics that handles WhatsApp conversations, clinic knowledge, leads, appointments, and human handoff.

DermoAI is a clinic-focused AI assistant designed to automate repetitive patient conversations while keeping important business decisions under application control.

Core workflow

WhatsApp → AI → Knowledge → Lead → Appointment → Human Handoff

Why DermoAI?

Clinics receive repetitive enquiries about:

• Treatments and services
• Consultation fees
• Doctors
• Working hours
• Clinic location
• Appointment availability
• Cancellations and rescheduling
• General FAQs

Handling these conversations manually creates repetitive work, delayed responses, and avoidable booking mistakes.

DermoAI automates the repetitive workflow while keeping staff involved when human judgment is needed.

How It Works

Patient → WhatsApp → DermoAI → Knowledge / Tools → Lead / Appointment
                                      ↓
                                 Human Handoff

The application controls business state such as appointments, leads, availability, and conversation mode. The LLM does not receive unrestricted database access.

Architecture

DermoAI uses a modular monolith designed for a single-clinic deployment.

WhatsApp
   ↓
Webhook
   ↓
PostgreSQL + pgvector
   ↓
Redis + BullMQ
   ↓
Conversation Engine
   ↓
RAG + LLM + Tools
   ↓
WhatsApp

The architecture keeps business rules inside application services while AI handles language, classification, retrieval, and response generation.

AI Pipeline

Message
  ↓
Intent
  ↓
Safety
  ↓
Conversation State
  ↓
RAG / Clinic Context
  ↓
Tool Validation
  ↓
Response

RAG

Clinic-approved information is stored and retrieved with PostgreSQL + pgvector.

Clinic Content → Chunk → Embed → pgvector → Retrieve → LLM

The system is designed to use approved clinic knowledge and provide a safe fallback or human escalation when reliable information is unavailable.

Appointment Booking

Appointments are handled by deterministic application logic rather than free-form AI decisions.

Request → Validate → Check Availability → Check Conflict → Commit

The server validates availability again during booking so the displayed slot is not treated as a guaranteed reservation.

Human Handoff & Safety

A conversation can move between:

AI ↔ HUMAN_TAKEOVER

Handoff can occur when:

• The patient requests a human
• Medical input requires staff involvement
• The AI cannot answer reliably
• An emergency signal is detected

DermoAI is not a diagnostic or prescription system.

WhatsApp

The inbound webhook flow is:

Verify → Deduplicate → Persist → Queue → Process

Provider message IDs are used for idempotency so duplicate webhook deliveries do not create duplicate processing.

Core Features

• WhatsApp patient conversations
• Clinic-specific knowledge retrieval
• AI intent detection
• Medical safety boundaries
• Lead management
• Appointment availability and booking
• Cancellation and rescheduling
• Human receptionist takeover
• Razorpay payment integration
• Knowledge management
• Clinic dashboard
• Audit logging
• Operational monitoring

Tech Stack

Layer	Technology
Frontend	Next.js, React
Backend	Node.js, Express.js
Database	PostgreSQL
Vector Search	pgvector
Background Jobs	Redis, BullMQ
AI	Gemini, LangChain
Messaging	WhatsApp Cloud API
Payments	Razorpay
Validation	Zod
Authentication	JWT, bcrypt
Styling	Tailwind CSS

Project Structure

dermoai/
├── apps/
│   ├── api/
│   └── web/
├── packages/
│   ├── schemas/
│   └── types/
├── docs/
└── package.json

Testing

The project includes tests for:

• Authentication and authorization
• Webhook verification
• Message idempotency
• Appointment availability
• Booking conflicts
• Conversation state
• Safety rules
• RAG retrieval
• AI tool execution

The project also includes a versioned AI evaluation suite for supported clinic conversation scenarios.

Security

Security-sensitive areas include:

• Staff authentication and authorization
• WhatsApp webhook verification
• Provider credentials
• AI credentials
• Database credentials
• Audit logging
• Clinic data boundaries

Secrets must never be committed to the repository or exposed in normal responses and logs.

Project Status

Active Development

Current focus:

WhatsApp → AI → Knowledge → Lead → Appointment → Payment → Human Handoff → Dashboard

AI handles conversation. The application controls business decisions.