BizCoPilot

The problem

Small shops in Myanmar sell almost entirely through chat. One customer asks about a jacket on Messenger, another sends a payment screenshot on Telegram, someone walks in and buys the last one, and the owner pieces it all together by scrolling. Orders get lost in the inbox, stock counts drift, and evenings go to answering the same price and size questions.

What I built

BizCoPilot gathers every channel into one order inbox and puts an AI assistant on the shop’s chats. The assistant answers customers in Burmese or English, searches the catalogue (even from a photo), builds a cart and takes the order through checkout. I built it with one other engineer. These are the parts I owned.

The bot’s first version. The Gemini integration, the tool-calling loop that lets the model search products, manage a cart and check out, and the Telegram bot.

Checkout in chat. Customers rarely send a clean address. The bot collects the address and phone across several messages, uses the model to check that an address really is one, reads Burmese numerals in phone numbers, and keeps every reply and button in the customer’s language.

Knowing what the AI costs. Every model call records its tokens, latency and cost, so pricing and model choices come from measured spend rather than guesses. Shops pay one credit per reply that actually reaches a customer.

The console that runs the business. A separate staff login with role-based access; a payment queue where credits are granted against real payments; refunds and adjustments that can’t run twice when a request is retried; and an audit record written in the same transaction as every action that moves money. A renewal list ranks shops by how many days of credit they have left, and a profit-and-loss page counts revenue as credits are spent, not when cash arrives.

Keeping the bots up. The backend counts what every bot sends and receives, alerts the team on Telegram when the bots go quiet, and has an outage switch that tells customers the shop will reply soon.

The marketing site in English and Burmese, the Docker build and the AWS deployment.

Outcome

• Live in English and Burmese on the web, on Android, and as an installable app on iPhone.
• Owners handle Messenger, Telegram and walk-in sales from one inbox, with the AI answering routine questions.
• The team runs payments, renewals and outages from a console where every action that moves money is audited.