Get Me to a Human

1. You

   Get me to American Express Purchase Protection about a stolen item.

2. Agent

   Calls American Express and listens to the phone tree.

3. Company

   “For Purchase Protection, press 2.”

4. Agent

   Presses 2, explains the handoff, and waits on hold.

5. Company

   “Purchase Protection, how can I help?”

6. Agent

   Calls you, confirms you answered, and joins both calls.

Recorded example. It does not place a call.

How it works

Every time someone speaks, the agent runs one short voice-processing loop.

1. 1
   Receive phone audioTwilio streams the live call to the MacBook over a WebSocket.
   Twilio Media Streams
2. 2
   Turn speech into textParakeet transcribes the latest audio locally so the agent can read it.
   Parakeet STT · local
3. 3
   Decide what to doThe conversation model reads the transcript and objective, then chooses a reply or an action such as pressing a key.
   DeepSeek · agent logic
4. 4
   Turn the reply into speechQwen generates the response audio on the MacBook.
   Qwen3-TTS · local
5. 5
   Send it back to the callTwilio plays the response or sends the selected keypad input. The loop repeats for the next turn.
   Streaming audio · DTMF
6. ✓
   Connect you when a human answersThe agent calls your phone and bridges both call legs. You take over with the representative already on the line.
   Human handoff