Skip to Main Content

DarFlow Agent

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

DarFlow Agent — Demo video

DarFlow Agent — an AI agent that runs the WhatsApp front desk and the expense back-office of a small Moroccan riad, built on Meta's Muse models. Pre-existing: DarFlow, my riad management software (closed source, not in the repo). Built this week: DarFlow Agent (everything in the repo), its WhatsApp agents and the small importer added to DarFlow (copied in darflow-importer/ for review). Timeline in the README; asked the organizers in #questions. The problem: Morocco has thousands of riads run by one manager and a few staff. Guests write on WhatsApp in every language at every hour, and expenses arrive as paper bills or are said out loud in Darija. I'm the manager of a riad: this is my daily life. DarFlow Agent handles both on one WhatsApp number: • Guest concierge: answers in the guest's language (Spanish, German, French, English, Darija in Latin letters), books dinner, records arrival times and transfers, resends the check-in link through tool calls, handles voice notes, and hands off prices, discounts and complaints to the manager instead of inventing answers. • Expense back-office: staff send a photo or a Darija voice note. SAM 3.1 finds and crops every sheet of paper in the photo, Muse Spark reads each crop (strict JSON schema) and re-checks the arithmetic and VAT, a deterministic policy flags duplicates, quotes, fakes and revenue documents. The manager replies "OK" and the expense lands in DarFlow, the riad's real software, through its own business rules. • Morning brief (added 1 Nov): every day at 08:00 the agent reviews the riad's day on its own, sends the manager one WhatsApp message (urgent / today / good to know) and proposes actions; the manager answers "1" or "tout" and the agent writes to each guest in their language. Facts and actions are built by code; Muse Spark ranks and phrases them; code checks every line against its fact. Design principle: the model reads and talks; the code enforces (3 h dinner notice, a guest only sees their own booking, fakes and quotes are never booked). Results: guest agent 21/21 live scenarios (incl. prompt injection, privacy probes, cancellation and complaint hand-offs, Darija in Arabic script); morning brief 3/3 live runs with 0 lines needing correction; documents 6/6 on real phone photos (incl. catching a supplier's VAT error); guardrail tests 16/16; DarFlow's full suite passes with the importer (248 tests).

Team