Building a voice agent that survives in production
Most voice agents sound great in a demo and fall apart on the first real call. A caller talks over it, changes their mind mid-sentence, switches languages, or pushes back, and the agent that looked flawless on stage starts making things up or loops forever.
Building a voice agent is fast now. Building one that holds up on real calls is the hard part. This hands-on session is about that second half.
We'll build a working voice agent, then put it through what production actually throws at it: 1,000 simulated calls across different personas and edge cases, scored on the failures a demo never surfaces (interruptions, objection handling, escalation, loops, transcription and speech accuracy). You'll watch it break, see exactly where, and fix it before it ever picks up a live call.
What you'll do
- Build a working voice agent (bring any stack, we'll have a starter template ready)
- Run 1,000 simulated adversarial callers against it (personas x scenarios)
- Score it on production-grade voice evals and find where it fails
- Harden it and re-run, using simulation & evaluation stack
Who it's for
AI engineers, founders building voice agents (support, sales, collections, BFSI), and PMs who own a voice product and need to know it's good enough to ship.
Bring
A laptop. We build and test live, so come ready to follow along.