Work / AI Kotib — Restaurant Secretary in Telegram
AI Kotib — Restaurant Secretary in Telegram
An AI secretary inside the Restaurant ERP: reports to the manager in Telegram at 08:30 and 18:00, answers any staff question about the day, and applies changes the manager asks for — after a ✔.
Reports on a schedule; commands become prepared changes that wait for ✔.
Every morning at 08:30 the manager gets a message: yesterday's sales and profit, today's bookings and shifts, tasks that were left unfinished, and a short plan for the day. At 18:00 comes the summary: what sold, what changed in the system, what needs attention. Any employee with access can ask for the same report whenever they need it.
The secretary knows every change in the ERP — orders, stock, staff, bookings, tasks — because it reads the same database through tools, not from memory. When the manager writes "move Aziz to the Tuesday evening shift" or sends a voice note, the change is prepared and shown with ✔ / ✖; nothing is applied until the manager confirms.
08:30 plan, 18:00 summary
Say it, confirm it
Text, voice, photo, PDF
Every change is prepared first
The manager learns things too late and types too much
A restaurant manager starts the day without knowing what happened yesterday and ends it without a summary. The numbers exist in the ERP, but reading five screens every morning does not happen. Small changes — a shift, a phone number, a stop-listed dish — mean opening the panel, finding the record and editing it, usually from a phone in the kitchen.
Reports without AI tokens, changes only through confirmation
The scheduled reports are built from the database by code and only summarised by the model when someone asks a question about them — so the two daily messages cost no quota and never invent a number. Every fact in an answer comes from a tool; if the data is not there, the secretary says so.
Changing tools never write. They validate the request, create a pending action with a human-readable summary, and the manager sees ✔ / ✖ buttons in Telegram. On ✔ the action is re-validated and applied, with an audit record of who confirmed what. Passwords and other secrets never pass through the model: a reset is done by the employee typing a new password into the bot, which deletes the message immediately and stores only the hash.
Timer, agent, Telegram
A systemd timer calls the ERP every few minutes; at the configured times it builds the report for each branch and sends it to the manager and the staff who opted in. Questions and commands come through the bot's webhook: the message is acknowledged in under a second, the Gemini call runs in a background thread and edits a "writing…" message as the answer streams. Voice notes are transcribed with a cheap model first; photos are shrunk before upload.
The agent shares its core with the assistant on this site: the same prompt structure, tool registry, action queue and logging, with the restaurant's own tools plugged in — sales, stock, staff, bookings, tasks and reports.
Two messages a day instead of five screens
The manager reads two messages a day and asks the rest in plain language. Routine edits take one sentence and one tap. Because everything is logged, the owner can see who changed what through the secretary. Next steps: voice replies and a weekly analysis with charts.