DineMate

Restaurants take orders by phone and app, then retype them for the kitchen: slow, error-prone, and invisible to staff until someone transcribes them. An AI ordering system that takes money is also an attack surface.

• Four LangGraph agents split the work: one extracts order details, one handles changes and cancellations, one pushes confirmed orders to the kitchen, one lets an admin update the menu and prices live.
• Voice ordering runs Whisper for speech-to-text and Torch TTS for spoken replies, so a customer can order without touching a screen.
• Orders stay editable for ten minutes, then lock, so the kitchen never works from a moving target.
• Role-gated dashboards for customers, kitchen, support and admin, with analytics on revenue, top items and peak hours.

• Red-teamed across 8 attack categories mapped to the OWASP LLM Top 10 and MITRE ATLAS.
• The critical finding: the browser could tamper with prices. Fixed by recomputing every total on the server.
• A Llama Prompt Guard 2 classifier screens input.
• The remaining findings, including an IDOR gap, are documented as accepted risks in the published report, not hidden.

• Never trust a total the browser sends. Recompute it on the server.
• Red-team before launch, then publish what is still open.
• Give orders a short edit window and lock them after it.