Munshi: A WhatsApp Front Desk for Small Hotels

Munshi is an open-source AI-driven solution designed to automate booking management for small hotels in Pakistan that rely on WhatsApp. The system utilizes a fine-tuned Qwen3.5-4B model to parse booking requests in Roman Urdu, English, or Urdu script into structured data. By integrating Python-based rule engines for calendar arithmetic and counting, the system achieves high accuracy, matching the performance of larger models like Gemma 4. The workflow incorporates TabPFN for risk-based advance pricing and Temporal for managing durable room holds. This approach emphasizes a 'models read, code decides' architecture, ensuring that critical business logic—such as date calculations and room availability—is handled by deterministic code rather than LLMs. The project, which is MIT licensed, demonstrates how small, efficient models can outperform larger ones when paired with robust, domain-specific code, significantly reducing operational costs while maintaining high reliability for hotel owners.