Brak
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Brak turns one phone recording into data a business can operate with. A seller walks through their products or shelves and says the prices; Brak returns a catalog with studio photos, sales copy and stock counts. Nothing goes live on its own: every result lands in a review tray where a person sees the original frame, the cutout, the studio photos, the price the model heard and the units the seller said, and publishes with one tap. Under the hood, Brak is an engine of configurable flows built end to end on the Meta Model API: five phases, seven model steps, verifier agents between steps with a hard cap on extra calls, and a per-video cache so a repeated recording costs nearly nothing. Each project has its own flow (steps can be switched off, custom Spark blocks add user-defined fields, and an assistant proposes a flow from a one-line brief) and its own API key, so any application can be a Brak client. The demo store, Market Lane, consumes Brak only through that API. Two templates ship today: Catalog (video → products with studio photos and copy) and Inventory (video → unit counts reconciled against what the seller said). Measured cost: a one-video, three-product run takes about 4 minutes and costs under US$0.50. Target users are the small shops of Latin America, where catalog and inventory still live in the owner's head. Built from scratch during the event (first commit Oct 29, 2026).