Dermoai solves this by acting as an autonomous, always-on digital receptionist and patient coordinator. Operating directly over WhatsApp—the primary messaging channel for outpatient clinics worldwide—Dermo greets patients instantly, answers treatment and pricing questions from verified clinic knowledge, checks live physician calendars, collects reservation deposits via integrated payment links, and intelligently escalates complex or clinical inquiries to human staff.
- Native integration with the Meta WhatsApp Cloud API.
- Handles multimedia, quick reply buttons, list messages, and natural conversational text.
- Zero delay: Responds in under 2.5 seconds with warm, brand-tailored conversational tone.
- Embeds and semantically indexes verified clinic documentation: services, doctor credentials, contraindications, preparation tips, and fee schedules.
- Uses semantic distance thresholding: if confidence is below safety boundaries, the system transparently defers to clinic staff rather than guessing.
- Strict two-phase reservation pipeline (
Check Availability$\rightarrow$ Soft Hold$\rightarrow$ Payment Commit). - Enforces doctor shift times, room constraints, buffer windows, and procedure durations.
- Completely immune to LLM hallucination: slots are fetched directly from PostgreSQL calendar tables.
- Automated Razorpay integration creates secure payment links inside the WhatsApp conversation.
- Collects consultation deposits (e.g., ₹500) to deter no-shows.
- Instant reconciliation via webhooks: slots are confirmed upon webhook verification.
- One-click toggle (
AI_MODE$\leftrightarrow$ HUMAN_TAKEOVER). - Receptionists can take over the conversation anytime from the web dashboard; the AI immediately pauses automated replies.
- Once the staff finishes, the conversation can be handed back to AI mode seamlessly.
- Built with Better Auth using server-side, HTTP-only secure cookie sessions.
- Email/Password credential authentication with robust password hashing and rate limiting.
- Social OAuth providers (Google, GitHub) pre-configured.
- Strict tenant isolation boundaries (
clinic_id) across all database queries.
WhatsApp delivers webhook notifications under an "at-least-once" guarantee. To prevent duplicate replies or double bookings, Dermo indexes incoming message_id hashes in a fast Redis cache with a TTL of 24 hours. Duplicate deliveries are immediately acknowledged (200 OK) and discarded.
The Large Language Model is strictly treated as an intent interpreter. When a patient says "Book Dr. Priya at 11am":
- The LLM extracts the parameter schema
{ doctor: "Dr. Priya", time: "11:00", date: "2026-10-01" }. - The schema is validated against a Zod validator.
- The booking service executes an ACID transaction on PostgreSQL with
SELECT ... FOR UPDATErow-level locks. - The LLM receives the outcome string and converts it into conversational confirmation.
Clinic documents (services, post-care instructions, pricing, doctor bios) are pre-chunked with overlap and converted into dense vector embeddings. Search queries execute hybrid cosine-similarity queries through pgvector indexed via HNSW (Hierarchical Navigable Small World) for sub-10ms retrieval latency.
Sultan
Full-Stack Engineer & AI Systems Developer
- GitHub: @sultanxdev