soko-pilot
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

In Kenya, WhatsApp is where many small-shop customers already ask their questions: “Is it in stock?”, “Which size fits?”, “What’s the price?”, “Can I send a photo?” Small merchants often can’t answer fast enough, and slow or repeated replies interrupt buying conversations. General-purpose AI chat isn’t a safe fix, because a model that guesses at prices or stock can damage a shop’s trust with its customers. SokoPilot connects to a merchant’s WhatsApp Business (Cloud API) number and answers customer questions in the chat, in text or from a photo, using a Meta model. Every product claim is grounded in the merchant’s own catalogue through scoped lookup tools, and the model never sets price, stock or payment truth. When details are missing, it asks a clarifying question. When it’s uncertain or the customer wants a person, it hands the conversation to the merchant, which pauses the AI until the merchant resumes it. Merchant control is built in. Merchants edit shop-specific instructions, preview replies before going live, and manage handoffs and takeovers from a single inbox that shows conversation history, customer images and replies. On the server side, SokoPilot checks merchant and customer ownership, deduplicates repeated webhook events, and honours STOP opt-outs. The order management and payment link generation tools still need stress testing and further evaluation, so are currently disabled, so the merchant handles the next steps after customer confirmation. The sales recovery functionality shows great potential in preventing lead drop-offs. Current status: a working prototype. We ran a versioned 24-scenario evaluation against the live model, and 20 passed (catalogue grounding, clarification, handoff, prompt-injection resistance, payment-boundary refusals, Kiswahili price and stock queries). We’ve noted the 4 failures and our latency (p50 8.6s, p95 13.3s against a 10s target), and are fixing them. A Sales Recovery feature, which flags stalled conversations for merchant-approved follow-up, is code-complete and tested but off by default. Next step: pilot with 10-20 SMB merchants in East Africa, measuring time-to-first-correct-reply