Voice AI ordering infrastructure

Rexi

Voice AI without giving the model control.

Call then VAD then STT then Interpreter then FSM then Pricing then Confirm then POS01Call02VAD03STT04Interpreterproposal05FSM06Pricing07Confirm08POS
  1. 01 Call
  2. 02 VAD
  3. 03 STT
  1. 04 Interpreter (proposal)
  2. 05 FSM
  3. 06 Pricing
  4. 07 Confirm
  5. 08 POS

The problem

LLMs are good at language. They are bad places to store business truth.

The boundary

The model reads the caller and proposes an intent. The server owns everything that can be wrong in a way that costs money: menu validation, price calculation, cart mutation, quote generation and confirmation.

Back to all work