Aahwan IT Group

Pizza Ghar: Multimodal Voice AI for Real-World Restaurants We didn't build a hackathon mockup. We built a production-ready Voice AI system for a LIVE restaurant client in Nepal. The Problem: Small shops in Nepal run on WhatsApp, but taking orders manually is chaotic. Generic AI chatbots fail on Nepali accents and Hinglish code-switching. The Solution: Pizza Ghar is an agentic voice ordering system built for PizzaGhar (Birtamode), running on top of our existing SaaS platform, Zappertrip (Aahwan IT Group, est. 2017), which already powers 5 live shops and 116 active customers. The Meta AI Integration (3 Models): Meta Muse Spark 1.3 (Text): Acts as the brain, fuzzy-matching messy transcription to our live MySQL menu and extracting strict JSON. Meta Muse Image 1.0 (Vision): Generates a professional, combined photo of the exact cart items asynchronously while the AI speaks. Meta Muse Voice Transcribe (Fallback STT): Integrated as a high-availability backup for speech-to-text. Technical Depth (Zero Hallucinations by Design): We engineered two distinct agentic personas with risk-appropriate autonomy: Customer Agent (Deterministic): The LLM only extracts JSON. A strict Node.js state machine controls the flow, guaranteeing zero AI loops or hallucinated orders. Owner Agent (Autonomous Text-to-SQL): The shop owner asks natural language questions (e.g., "Which items are low on stock?"). Muse Spark writes a live MySQL query (using JSON_TABLE for nested data), executes it against production, and returns the answer. We enforce strict SELECT-only guardrails. The Closed Loop: The system handles the full lifecycle: Voice Order -> AI Image Generation -> Owner Dashboard -> Payment Page -> Inventory Auto-Deduction. Live Links: Platform: https://zappertrip.com Hackathon Demo: https://whatsapp.zappertrip.com/hackathon/ GitHub: https://github.com/sitrsh12/meta-ai-hackathon-Aahwan_IT_Group