Bazaar-AI: Zero-Cost WhatsApp AI Employee

Bazaar-AI is a zero-cost AI employee for small merchants in underserved-language markets, built for Track 3: LocalFirst. It talks to customers and shop owners in English, Hindi, Hinglish, Tagalog, or Taglish over a direct API or straight through WhatsApp, handling everything from quick product questions to fully negotiated bulk orders with auto-generated invoices. The core idea is a ComputeRouter: a semantic classifier that scores every incoming message on negotiation intent, invoice/tool-calling signals, code-switched vernacular density, and conversation history, then routes it to one of two tiers, a fast low-VRAM edge model for simple lookups, or a deep-reasoning heavy model for negotiations and structured invoice extraction. A direct API endpoint and a Twilio WhatsApp webhook both run through the exact same routing pipeline, so behavior never drifts between entry points. The architecture targets self-hosted Gemma-4 models on AMD Instinct MI300X GPUs via vLLM. During public evaluation, a transparently-disclosed Model ID Translation Layer substitutes free Fireworks serverless models with equivalent parameter counts, so the routing logic can be demonstrated at zero API cost without dedicated GPU hosting. Moving to the real Gemma-4 deployment needs only an endpoint config change. A live Streamlit dashboard gives judges a real-time view of routing decisions, compute saved, and cost ($0.00 throughout). A 17-check automated regression suite validates routing precision, webhook handling, and invoice JSON extraction. Phases 1 through 3 are complete and verified end-to-end, with a hardening roadmap (guided decoding, ML-based vernacular detection, persistent storage, webhook signature verification) scoped for post-hackathon work.