TrekDesk AI - AI-Powered Tour Booking Assistant
The Narrative
Tour operators often struggle with handling real-time inquiries and bookings across different time zones, leading to missed opportunities and slow response times.
I built an AI-driven assistant that can handle complex tour inquiries, check real-time availability via Google Calendar, and process bookings autonomously using tool-calling and RAG.
Key Features
Real-time Voice & Chat
Implemented low-latency communication using WebSockets and Gemini Live API for a natural human-like interaction.
RAG Pipeline
Built a semantic search engine using PostgreSQL and pgvector to provide accurate information based on private tour data.
Automated Scheduling
Integrated Google Calendar API for dynamic availability checks and automated booking confirmations.
Engineering
// SYSTEM_OVERVIEW
Real-time AI voice system with RAG capabilities, containerized and deployed on Google Cloud Run.
01_MODULE: VOICE-FIRST AI ARCHITECTURE
The system uses a high-performance Node.js backend to bridge the gap between frontend WebSockets and the Gemini Live API, ensuring sub-second response times for voice interactions.
02_MODULE: EMBEDDABLE WIDGET SYSTEM
Designed a secure, domain-validated widget system that allows third-party websites to integrate the AI assistant with just a few lines of code, featuring full session persistence and analytics.
Results
Live product handling tour inquiries through real-time voice and chat.
Deployed on Google Cloud Run with an embeddable, domain-validated widget for third-party sites.
Automates availability checks and booking confirmations via Google Calendar integration.
TODO: add verified metrics (response latency, inquiries handled, bookings completed) once measured.