Dastan

SaudaAI is an AI assistant that works on top of the shop a small seller already has. Millions of small shops in Bangladesh sell through a Facebook Page and a simple online store, run by one or two people from a phone. The owner answers the same questions all day (price, size, stock, delivery charge), often by voice and in Banglish. Orders are copied by hand, totals are added in a notebook, stock drifts from what the website shows, and payment reminders are forgotten. SaudaAI connects to the seller's existing store (Shopify, WooCommerce, or any site through a small Store API) and to the places customers chat: Messenger, Instagram, WhatsApp and a chat bubble on the shop's own website. It answers customers in Bangla, Banglish or English with live prices and stock from the store. It sends the store's product photos, takes the order details and creates the order in the store. It sends the confirmation, the invoice and a payment link. The owner runs the shop from Telegram or the dashboard by text, voice note or photo: - "Aj koto sell holo?" (how much did we sell today?) gets today's sales. - A phone photo with a price becomes a clean product photo and a new product in the store. - "Make a winter ad and post it on the Page" gives ready ads with Bangla text, posted after one tap. The owner stays in charge. Anything that commits the shop or touches money waits for one tap: confirming an order, payments, cancellations, discounts, new products and posts. Prices and totals are always computed by code, never written by the model. Replies that mention a number not found in the shop's data are held for the owner, and every step the assistant takes is shown in an agent log. An analytics page shows sales by channel, how many chats the AI handled, and what customers asked for that the shop does not stock. Meta models: - Muse Spark is the agent: tool calling and reading product photos. - Muse Voice Transcribe turns Bangla and English voice notes into text. - SAM 3.1 cuts the product out of a phone photo. - Muse Image builds ad scenes around the real product photo. The Bangla text and prices on the ads are drawn by code. An evaluation suite of 51 real-model scenarios covers ordering, stock, payments, safety (prompt injection, discount pressure, a customer trying owner actions) and connected stores. It passes 51 of 51, both on the built-in catalog and with a connected store.