Md Amir Hossain

Independent resorts with no night front desk lose after-hours guest enquiries — and a wrong answer (wrong price, invented availability) is worse than silence. Vivereply Concierge is a multi-tenant AI front-of-house: each resort gets its own rooms, booking backend, knowledge, voice, staff scope, and logins. On the live demo (Krishnachura Retreat, Himchari, Cox's Bazar), a guest chats on the resort website: the agent checks live availability, quotes exact prices from catalogue rates, holds a real booking draft with a code (PENDING hold, Suite total governs), collects payment through the resort's own pay page (booking deep link the model cannot invent), verifies bookings privately (dates + first name + status only; email/phone lookups get one generic reply), answers in Bengali or English, and hands the impossible to staff with the full conversation on a triage board. Muse Spark (muse-spark-1.3 via the Meta Model API) is the decision-maker: every turn runs a hand-rolled tool-use loop (max 6 rounds) over room-scoped tools built per request from the property's own catalogue. Token usage per backend is logged on every call; one Baseten-hosted fallback covers primary outages. RAG is supporting-only (threshold 0.45, top-4, silence over guessing) — never the product. First commit Oct 27, 2026 — inside the Oct 26–Nov 4 window. Eval 19/19, probes 10/10, 98/98 unit tests. Solo founder. Live: https://app.vivereply.com — Demo resort: https://demo1.vivereply.com