A sales agent that speaks Bangla at three in the morning
A multi tenant WhatsApp assistant for small businesses, built to sound like the owner and to know when to stop talking.
- AI assistant
- Messenger
- Telegram
- Self hosted
Small businesses in Bangladesh do not lose customers because their product is wrong. They lose them because the reply came eleven hours late, while a competitor answered in eleven seconds.
Hiring someone to sit on a phone overnight is not a business model. So the shift went to software, with one condition attached: it had to be better than silence, not just faster than it.
Read the rest
What makes Bangla hard
Most assistants are demonstrated in tidy English, one question per message. Real conversations here are nothing like that.
People send voice notes, because typing Bangla on a phone is slow. They split one question across four messages and expect all four to be remembered. They switch scripts mid sentence, writing Bangla in Latin letters and back again. They ask for the price, then the delivery charge, then whether the price included the delivery charge, which is the same question wearing a hat.
The assistant reads voice notes. It waits before answering, because the next message is usually still coming. It answers from the business's own catalogue and the business's own prices, not from a general impression of what such a product might cost.
The part that sells it
When the assistant does not know, it says so and fetches a human. That single behaviour is what makes owners trust it enough to leave it running.
An assistant that guesses will eventually quote a price nobody agreed to, promise a delivery date nobody can meet, or argue with a customer who is already angry. Every one of those is more expensive than a slow reply. So the handover rules are strict: money disputes, complaints, anything outside the catalogue, and the conversation goes to a person, with the history attached so nobody has to ask the customer to start again.
Why it runs on our own machine
The obvious way to build this is to rent everything and bill per message. That works beautifully while the business is small and punishes it precisely when it starts working.
Instead the whole stack sits on a server we own. Automation, database, chat and the model gateway, side by side, one machine, a bill that does not grow every time a customer says thank you.
The demo number is live. It is on the front page of this site, and it will answer you right now, whatever time you are reading this.