START: Alex Kolchinski, Founder & CEO, ThunderPhone "Voice AI that nails the calls others fumble"

Listen on Apple Podcasts Listen on Spotify Watch on YouTube Subscribe to START

‍

‍

AI coding agents learned how to build voice AI from the internet. That might be the problem.

The standard approach is simple. Transcribe the caller. Send the text to a model. Turn the answer back into speech.

Three steps. Countless posts and repositories. Repeated so often that coding agents tend to reproduce the same architecture.

Alex Kolchinski's bet at ThunderPhone is that the default is already becoming obsolete.

Instead of one fixed pipeline, his team mixes and matches models. Sometimes several cheaper ones instead of one expensive one. Sometimes audio goes straight to the LLM, skipping transcription entirely.

Every decision comes down to three competing priorities: quality, speed, and cost.

Alex calls it the Ferrari approach. While other teams have good reasons to build systems that work well enough to scale, ThunderPhone focuses on squeezing more performance out of every component.

And there's a catch. Alex doesn't expect that technical advantage to last forever. Foundation models keep improving, and coding agents are beginning to figure out better architectures themselves.

Until then, the edge is knowing what to build differently.

ThunderPhone: "Voice AI that nails the calls others fumble"

‍

🎙️Alex Kolchinski, Founder & CEO, ThunderPhone on Fondo START  

‍

01:07 Why ThunderPhone moves beyond the standard voice AI pipeline

01:42 Navigating the tradeoff between quality, latency, and price

02:35 Why Storm prioritizes getting real customer calls right

03:45 How software and B2B changed Alex's founder playbook

04:45 Why the traditional three-step voice AI pipeline is becoming obsolete

05:00 How voice AI is outperforming humans on transactional calls

06:32 Why ThunderPhone isn't training its own foundation model

08:15 How two founders built the platform in 6–12 months and where coding agents fall short

09:30 The Model T versus Ferrari approach to building voice AI

10:15 Spark at 2¢, Bolt at 5¢, and Storm at 9¢ per minute

‍
Check out thunderphone.com