Sarvan

THE PROBLEM Small businesses get plenty of engagement on Meta's platforms but few sales. Comments asking "price?" go unanswered, DMs wait for hours, and when it gets busy the owner loses track of who asked what and who was ready to buy. Every unanswered comment is a lost sale. WHAT SALESIE DOES - A shopper comments on a post. Within seconds Salesie replies privately in Messenger with the real price from the catalog. - The agent qualifies them (colour, size, where to ship), saves it on the deal, and sends a checkout link when they're ready. - Payment marks the deal won and sends a receipt. Every step moves the deal through the pipeline: new, engaged, qualified, payment sent, won. - Refunds, complaints or questions the catalog can't answer are handed to the owner ("Needs you"). The owner can take over any chat; the agent waits until resumed. - The dashboard shows the pipeline, every conversation, and which posts actually made money. LIVE ON A REAL FACEBOOK PAGE Salesie runs Bloom & Thread, a demo fashion shop on Facebook. A real "price" comment became a private reply, a chat about size and colour, a checkout link, a test payment and a won ₹2,499 deal, all handled by Muse Spark with no human involved. HOW IT USES META'S AI - Muse Spark 1.3 (Meta Model API) is the sales agent, working through tool calling: catalog search, product details, saving the shopper's choices, deal stages, payment links and hand-off to the owner. Prices only come from the catalog, so the model can't invent a discount. A "Why?" panel shows every tool call behind each reply. - Muse Image 1.0 made the shop's brand: the Page's logo and cover, six product photos and three offer posts, using the edits endpoint to keep products identical across images. - Meta's platform: Page webhooks for comments, private replies, Messenger webhooks, and message echoes to detect when the owner replies. BUILT FOR REAL USE - Shopper text is data, never instructions, and only Salesie's own checkout links can leave in a reply. - Someone who messaged first and commented later is merged into one contact. - Signed, idempotent webhooks; per-shopper locking; limits on model spend; a cut-off model reply is never sent. - 336 integration tests against a real Postgres database. Stack: FastAPI, PostgreSQL, React. BUILT DURING THE HACKATHON Everything in the repository was written during the hackathon (first commit Oct 29): the CRM, webhook pipeline, agent, dashboard, Facebook and Messenger integration, and the Bloom & Thread brand. WHAT'S NEXT Connecting Instagram (built and tested in Salesie's simulator), WhatsApp with Muse Voice Transcribe for voice notes, and SAM 3.1 with Muse Spark vision to recognise products in new posts.