A café's sales, costs, and stock, logged from Telegram.

One person maintains this, the staff already had Telegram open all day, and the café had no budget for a subscription. Those three constraints decided almost everything about how it is built.

• No serverThe whole thing runs in Apps Script on the free quota, against the café's own Google Sheet. Nothing to host, nothing to renew.
• No new appOrders are typed the way staff already speak. There is no screen to learn and no barcode to scan.
• No redeployItems, prices, and recipes are rows in a sheet, so a menu change is a cell edit.
• No silent guessingThe model may only return items that already exist on the menu, and a wrong receipt line is fixed with one follow-up sentence.

Everything runs in Google Apps Script, so there is no server to host or pay for.

1. Poll or pushA time trigger polls Telegram for fifty seconds at a time, then hands back to the trigger. The offset is saved in Script Properties so nothing is logged twice. A webhook mode does the same job without the wait.
2. ParseA typed line splits into quantity, item, and size, then gets priced from the Menu sheet. Sentences that do not fit, like "add kani xl, kani m to my sales", go to the model, which may only answer with items already on the menu.
3. LogSales, expenses, and inventory live in separate tabs. Each sale deducts ingredients and warns when something drops below its minimum.
4. Read receiptsPhotos go to a vision model that returns JSON. Personal items are dropped, shop supplies are logged and restocked, and one follow-up sentence corrects a wrong call.

Staff send these as plain messages. Anything else goes to the model as a question, and menu changes need no redeploy.

• 1 latte 16log an order
• add kani xl, kani m to my salesseveral at once
• expense 588 beanslog a cost
• recipe spanish lattedrink build
• stockinventory levels
• todayweekmonthsales
• besttop sellers today
• profit todayrevenue minus costs
• expenses todaycost total
• menu addmenu listmenu deleteedit the menu
• undoremove the last sale
• undo expenseremove the last cost

Four excerpts. The full script stays private because a real café runs on it.

var PROPS = PropertiesService.getScriptProperties();
var TOKEN = PROPS.getProperty("TELEGRAM_TOKEN");
var SHEET_ID = PROPS.getProperty("SHEET_ID");
var OPENROUTER_API_KEY = PROPS.getProperty("OPENROUTER_API_KEY");

function parseOrder(text) {
  var lower = text.toLowerCase().trim();
  var match = lower.match(/^(\d+)\s+(.+?)(?:\s+(\S+))?$/);
  if (!match) {
    return { error: 'Format: "1 latte 16" or "2 kani-mango s"' };
  }
  return {
    qty: parseInt(match[1], 10),
    item: match[2].trim(),
    size: match[3] || "16",
    error: null
  };
}

var prompt = "Parse the user's message into a JSON list of orders.\n" +
  "Available menu items (name (size)): " + menuList.join(", ") + "\n" +
  "Return ONLY valid JSON: " +
  "{\"orders\":[{\"qty\":number,\"item\":\"menu item\",\"size\":\"size\"}]}\n" +
  "Only include items that exist in the available menu list.";

var payload = {
  model: "minimax/minimax-m3.1-flash:free",
  response_format: { type: "json_object" },
  temperature: 0,
  messages: [{ role: "user", content: prompt }]
};

var start = textResponse.indexOf('{');
var end = textResponse.lastIndexOf('}');

if (start !== -1 && end > start) {
  textResponse = textResponse.substring(start, end + 1);
} else {
  textResponse = textResponse
    .replace(/```json/g, "").replace(/```/g, "").trim();
}

var resultJson = JSON.parse(textResponse);

What it does not do, written down so nobody assumes otherwise.

• Polling has a waitThe time trigger holds the connection for fifty seconds, so a reply can lag a poll behind. Webhook mode removes the wait but needs a public endpoint.
• The model can still misreadIt is constrained to the menu, not guaranteed correct. Receipt photos need a readable total line; a folded, dim receipt gets a correction request instead of a guess.
• One café, one SheetTwo branches would share the same tabs. It is a small-shop tool, not a point-of-sale system.
• No reportsNumbers land in Sheets already, so anything longer than a daily total is done there rather than in the bot.