🥞 Try DosaDash

1 · Log in (any phone works)

Signup is OTP-based and runs in demo channel: enter any Indian mobile number (e.g. a made-up 98765 43210) and the OTP is shown right on the screen — no SMS is sent. Each number becomes its own customer account, so you get a clean cart, order history and “my usual” memory of your own.

If you link the Telegram bot later, OTPs are DM’d there instead — that’s the swappable OtpChannel interface doing its job.

2 · The three surfaces

There is also a Telegram bot (@dosadash_bot) — same order agent, different adapter. Try sending it a voice note in English or Tamil.

3 · Test payments (Razorpay TEST mode)

After checkout, the pay button opens Razorpay’s test checkout. Use the official test credentials — no real charge is possible on a TEST key:

Real cards are rejected in test mode by design. If the environment runs without Razorpay keys, the button becomes a one-click demo payment instead (the PaymentProvider interface swaps providers without touching checkout).

4 · AI things worth trying

• Order by chat (💬 on the menu page): “2 masala dosa and a filter coffee” — then edit it: “make one of them onion dosa instead”. Try Hinglish (“ek plate idli bhi add karo”), Tanglish, or full Tamil (“ஒரு மசாலா தோசை”). Every item is validated against the DB — the agent cannot invent dishes.
• Ask about food: “is the ghee roast vegan?”, “what has peanuts?” — answers come from a RAG pipeline with citations.
• “My usual”: place the same order twice (on different days), then ask for “my usual”.
• Support (🛟 on the orders page): ask where your order is, or try to talk it into a refund — refunds always escalate to a human inbox.
• Menu photos: dishes with a ✨ badge are AI-generated images that went through owner approval — the label is permanent by design.
• Backoffice AI (as demo admin): the analytics copilot (“top 5 dishes by revenue last week” — read-only SQL with a guardrail), the eval scoreboard, LLM cost + cache dashboards, AI-drafted purchase orders on the Inventory tab, review sentiment scored by a local fine-tuned model.

5 · House rules