Your voice→Transcribe→Reason→Speak→Playback
Press Start call and allow the microphone. Try "what did he build at IntelePeer?", "how big a team has he run?", or "what kind of role is he looking for?"
- Transcribe —
- Reason —
- Speak —
Batch pipeline — each stage finishes before the next begins, so they add up rather than overlap. Typically four to five seconds end to end. Production voice starts transcribing while you are still talking and starts speaking before the model has finished writing, which is most of the distance between this number and a call that feels natural.
How it is grounded
The agent answers from a fixed knowledge file I wrote by hand. It is not searching the web and it is not reasoning freely about my career. If you ask something the file does not cover, it is instructed to say so and point you at my email rather than fill the gap.
That constraint is the whole point. An agent that improvises about someone's work history is worse than no agent, and the failure is invisible until someone repeats the invention in an interview. The prompt also refuses compensation questions and will not characterize decisions made by a former employer.
How it is built
Three model calls chained behind one endpoint, with the key held server-side in a Cloudflare Function. The browser sends audio and receives audio. It never sees a credential.
The engineering that matters is the constraints: a cap on audio length per turn, a cap on reply length so the speech bill cannot run away, conversation history trimmed to the last few turns, and a graceful path when speech synthesis fails so the turn returns as text instead of dying.
What separates this from production
The round trip you are seeing is what a batch pipeline costs. A deployed agent is architected differently, and the difference is worth naming precisely rather than hand-waving at it.
- No streaming. Each stage completes before the next begins, so the latencies add up. Production systems start transcribing while you are still talking and start speaking before the model has finished writing, which overlaps most of that cost away.
- No barge-in. This is push-to-talk. A real agent lets the caller interrupt mid-sentence, which needs bidirectional streaming and voice activity detection. This is the genuinely hard part of production voice, not the model choice.
- No telephony. Browser mic, not a carrier leg. Add SIP or WebRTC transport and an entire category of failure modes arrives with it.
- No scheduling. Ask it to book time and it will hand you my email. That is the next thing to build.
- Rate limited. It is a demo, not a service.