I built Dohos. A good conversation has to become a correct order.
Voice AI that takes restaurant phone orders, developed through Yale’s Tsai CITY startup program. It has handled about 2,000 real customer calls in live testing.
- Real calls
- About 2,000
- Stage
- Live testing
- Built in
- Rust and Python
- Program
- Yale Tsai CITY
It began with a phone I used to answer
I managed our family’s restaurant in Baltimore through college. I rebuilt phone-order coverage there with VOIP and offshore staff. That made restaurant calls a concrete place to start with voice AI.
The model proposes. Code decides.
A caller can say “actually, make that two” halfway through an order. The language model interprets the request and proposes a change using references to menu items. Application code checks the menu and the options, calculates the price and applies the change. Any material change means the caller confirms again.
Holding down the cost of every call
The menu and tool descriptions had grown too large, so I cut the model’s context. I also built custom voice-activity detection in Rust, with no runtime dependencies and tests that compare it against the reference model.
What I own
Architecture, model selection, the order flow and call review. I use coding agents to implement and check their work through tests and by listening to calls. The next test is what breaks when it runs at more locations.