Diwan

Diwan is an AI back office for a Dubai real-estate brokerage, built on Meta’s Muse models, where one sentence from a client updates the whole office, and nobody can act on what’s no longer true. In real estate, clients change their minds, and the change rarely reaches everyone. Finance still expects money from the wrong person, approvals stay valid, and Compliance hears last. In Diwan, every task, approval and draft rests on the client’s exact words. When Karim says “Actually, forget my brother. I’ll pay it all myself”, Compliance is told first, only the work that rested on the old statement is withdrawn, Finance’s approval goes back for review, and the old approval can no longer be executed. When the meeting ends, the work is handed out as tasks with an owner, a date and the words they came from. 64 chartered agents across ten rooms, watchable live on a 3D office floor: Rania answers clients on WhatsApp in English and Arabic from real listings only; Nova, the owner’s assistant, cites a record for every answer; the Studio writes scripts, makes pictures and pauses a post when a listing’s price changes. Meta models: Muse Spark runs the agents · Muse Spark Open on Baseten runs Compliance’s source-of-funds analysis and its devil’s advocate · Muse Voice Transcribe hears meetings with speaker labels · Muse Image makes the marketing pictures. Agents prepare; people sign. Every model call is logged, confidential compliance work is hidden from brokers, and nothing is sent to a client or published without a person’s yes. Built from scratch during the hackathon (27 Oct – 4 Nov 2026). Demo data is fictional. Try it: https://diwan-kxp4.onrender.com Sign in as: layla (owner: sees everything, including the 3D office) · password: DiwanOwner-Demo26