An AI concierge that handles restaurant and nightclub bookings over WhatsApp and voice: it checks availability, makes the booking and recommends venues through MCP server tools.
WhatsApp through WATI, and voice calls through RetellAI
2 channels
Availability, booking and confirmation in one conversation
Chat to booking
Guest conversations scored after every iteration
Evaluated
The starting point
The challenge
Guests of a luxury concierge expect instant, personal answers on the channels they already use, at any hour. Staff were answering the same availability, booking and recommendation requests on WhatsApp and the phone, and a website chatbot would never reach guests where they actually message.
The solution
What we built
Built the concierge agent in PydanticAI on GLM models from Z.ai. It understands the request, extracts the booking details and manages the conversation.
Connected WhatsApp through the WATI WhatsApp Business API, and phone calls through RetellAI voice agents.
Exposed availability, bookings, venue information, guest preferences and notifications to the agent as MCP server tools.
Grounded answers about menus, venues, FAQs and policies in a RAG knowledge base, with conversation memory for context and preferences.
Confirms each booking and sends the details back to the guest on WhatsApp.
Architecture
How the system works
How it fits together Data flow
01GuestWhatsApp message or phone call
02ChannelsWATI and RetellAI
03Concierge agentPydanticAI on GLM
04MCP toolsAvailability, bookings, venues
05KnowledgeRAG and conversation memory
06ConfirmationBooking details on WhatsApp
Engineering
How it was engineered
Built evaluation sets of guest conversations covering availability checks, new bookings, changes and cancellations, and recommendation requests.
Ran the evaluations after every iteration and against production conversations.
Traced every agent run and tool call in Logfire.
Scope
What was delivered
01
Concierge agent
02
WhatsApp channel through WATI
03
Voice channel through RetellAI
04
MCP tools for availability, bookings, venues, preferences and notifications
05
RAG knowledge base and conversation memory
06
Evaluation sets and Logfire tracing
Gallery
Screens and diagrams
From WhatsApp message to confirmed booking: the agent, its MCP servers and the knowledge layer.
An agent that works inside the channels pet parents already use: it answers Instagram and Facebook messages, qualifies the lead and books the appointment in MoeGo, keeping GoHighLevel up to date.
An AI agent layer for a diagnostics network whose patient app has 1M+ downloads: plain-language lab-report explanations, test guidance, lab lookup and voice, orchestrated across MCP tool servers.