DepaPranata

I work as a waiter in a busy restaurant in Medan, Indonesia. My shifts are 11 hours on my feet, and this problem is mine, firsthand: when guests don't speak our language — especially tourists from China, Korea, and Japan — orders go wrong. Explaining the menu in broken English while the floor fills up wastes the little time a short-staffed team has. And our house rules forbid waiters from pulling out a phone to translate on the spot. MejaBahasa solves it with a door that is safe under those rules: a QR code on the table opens a web chat on the guest's own phone — the waiter never touches theirs. Guests browse a visual menu (47 languages), add items to a cart, and check out. The AI detects their language automatically (zh / ko / ja / id, no manual picker) and asks about allergies, requests, and disliked ingredients before anything else — an allergen clash rejects the whole order and names the item. The agent then asks dine-in or takeaway; for dine-in it asks "got a table yet?" and the guest scans the table QR with the in-app camera — the table number is read from the QR itself, never typed. Guests without a table can take away or join a waiting list with a queue number, and see which tables are free. The agent is not a chatbot wrapper: it validates every order against the real menu and database, asks back in the guest's language when anything is ambiguous, and refuses to guess. Only after the guest confirms AND staff taps ACC does a clean Indonesian ticket reach the kitchen display (open → cooking → ready → served). Staff get a live table-occupancy board (green = free, red = occupied) plus waiting-list management. A floor-coordinator screen assigns cooking and delivery tasks to the lightest load, and the owner dashboard shows revenue, order statuses, top-5 items, tickets, low stock, and complaints. Every agent action is written to an audit log. One engine, one SQLite database: guest chat, staff dashboard, kitchen display, floor coordinator, and owner dashboard. 200 automated checks pass