Small service businesses like fitness studios and salons juggle bookings across messaging apps, calendars, and payment links. As volume increases, missed messages, double bookings, and manual follow-ups become the norm.

Customers want to book the way they already communicate - through WhatsApp. But building a conversational booking experience that handles natural language, manages capacity in real-time, and prevents race conditions is a complex engineering challenge.

We built a WhatsApp-first booking experience where customers book using natural language like "I want yoga tomorrow at 7pm". The AI layer uses LangGraph and LangChain with a ports and adapters architecture - keeping model, tool, and provider concerns isolated as the product evolves. Typed tools handle session discovery, booking management, crew info, locations, memberships, preferences, and business operations.

The NestJS backend uses MongoDB with atomic operations to ensure race-condition safe bookings. Real-time capacity management includes waitlist support with automatic promotion when spots open up. The multi-tenant architecture supports multiple businesses with complete data isolation.

• - Broad unit, integration, E2E, and eval coverage across the API, AI flows, and dashboard surfaces.
• - Per-channel latency budgets, tool timeouts, retries, and circuit breakers keep WhatsApp turns bounded.
• - Race-condition safe: 50 parallel booking requests on a 10-capacity session → exactly 10 succeed.
• - Redis-backed caching, MongoDB indexes, and an outbox-style notification path support operational reliability.
• - Multi-tenant architecture supporting multiple businesses with complete data isolation.