Hospital Appointment AI Agent

Problem

Hospital front-desk appointment booking over phone or chat is repetitive, staff-intensive, and error-prone — it needs an agent that can hold a genuine multi-turn conversation, track slot-filling (doctor, date, time), stay strictly on-topic, and take real actions rather than just talk.

Approach

• 01Built a LangGraph state-machine agent (persona 'তানিয়া ইসলাম') that tracks slot memory across turns — doctor selection, natural-language date/time parsing (dateparser), and domain-strict guardrails that keep the conversation on-topic.
• 02Added a dedicated Gemini Live bridge alongside the existing STT→LLM→TTS pipeline so the same appointment tools, prompts, and conversation state can drive either native Gemini Live audio or the classic loop — an interchangeable-backend design.
• 03Delivered both a WebRTC voice transport and a text/chat transport from one FastAPI backend, with Twilio call/SMS and email tool-calling for appointment confirmations.
• 04Added an internal RVC (voice-conversion) module and a lightweight DB-view/admin API for operational visibility into bookings.

Results & Impact

• Runs two interchangeable generation backends — native Gemini Live audio and a classic STT→LLM→TTS loop — behind one appointment-booking agent core.
• Takes real-world action (Twilio call/SMS and email confirmations), not just conversation, on top of a domain-strict, slot-memory LangGraph core.