FunnelChat Génesis

Génesis turns one WhatsApp voice note into a live WhatsApp business in under a minute. The problem: 54.7% of Colombia's workers are informal (DANE, GEIH), and of the 2.4 million micro-businesses in the 24 main cities only 17.4% have a computer, tablet or laptop (DANE, EMICRON 2025). Their customers already write to them on WhatsApp, but selling there properly takes a business profile, a catalog, an order form, payment links and someone answering at 6 a.m. All of that is set up from a computer they don't have. How it works: Rosa, who sells arepas in Cali, sends Génesis a voice note and two photos and types "listo". 1. Muse Voice Transcribe returns a timed transcript; Muse Spark 1.3 compiles it into a typed BusinessSpec (products, prices, hours, neighborhood, payments), with every field traced to the second of audio where she said it. 2. If one essential fact is missing, Génesis asks exactly one question, reads the summary back and waits for her "sí". 3. A plan/apply graph builds the business in parallel lanes, with idempotency keys and automatic revert of critical failures: business profile, SAM 3.1 product crops, a Muse Image logo and a clean catalog shot made from the crop, a WhatsApp Flow order form, a business-only Spark agent with tools, and Mercado Pago payment links. 4. Before launch an Inspector, Spark posing as a customer, audits the business: prices, hours, delivery, a beer it must refuse, a test order through the Flow and a test payment. The business only opens when the Inspector approves it. 5. Customers then ask, order in a native WhatsApp Flow and pay inside WhatsApp; Rosa gets every order and payment alert on her phone. In a real run with the live Meta models the whole thing, from "listo" to "your business is live", took 51.4 seconds, and the Inspector passed 9/9. Engineering: Python/FastAPI, a deterministic plan/apply engine with SQLite state, an event bus streaming every step to a live panel, a WhatsApp Cloud API client (Flows 7.3, interactive messages, resumable upload, per-number webhook override), signed webhooks, and 231 offline tests. Running Génesis against the real models exposed failure modes that mocks never showed (a re-compile that duplicated products, empty hours, SAM missing long product names); each one is now handled in code and covered by a regression test built from the recorded real outputs. Who benefits: the millions of informal sellers in Latin America who already live on WhatsApp but will never open a dashboard. Rosa doesn't learn a platform: she talks, and her business answers, takes payment and alerts her.