e-bot

Bangladeshi e-commerce lives on Facebook. Stores manage product catalogs, customer queries, and orders entirely through Messenger — but every reply is typed by hand. The same questions come in hundreds of times a day: price koto, XL ache, delivery charge koto? e-bot was built to answer all of them instantly, in the customer's own language, without a single human keystroke.

The backend is a FastAPI service that receives Facebook webhook events and queues them as Celery tasks. Each message goes through a two-step LLM pipeline: the model either calls a tool (product search, store info, order placement) or replies directly. Product images are embedded with DINOv2 and stored in PostgreSQL with pgvector — so when a customer sends a photo, the bot identifies the product by visual similarity. Redis caches the current product in focus for 5 minutes, making follow-up questions ("any discount?", "XL ache?") instant without re-hitting the database.

Placing an order through e-bot feels like talking to a polite shop clerk. When a customer signals intent — "I want one," "order korbo" — the bot confirms the product, asks for the delivery address and zone, summarises the order with totals, and waits for a 'yes' before pushing it to the store's order log. No forms, no app downloads, no callbacks to wait for. The entire transaction happens inline, in whatever language the customer is comfortable typing.

Language detection runs on every incoming message. If a customer types 'price koto?' in Banglish, e-bot answers in Banglish. If they switch to English mid-conversation, the bot switches too. For store owners, that means no English-only filter on their reach — the bot serves the customer in the dialect they actually use, which in Bangladesh's online market is often the difference between a sale and a scroll-past. Underneath the casual tone is a stricter system prompt that keeps the bot honest: it won't invent prices, won't promise stock it can't verify, and hands off to a human when the customer asks something outside its scope.