Founder Spotlight

It’s fun building a voice agent until every little modification becomes a mystery. Change the prompt, change the model, take one step in the call flow and suddenly the agent is better or worse or just acting strange for reasons no one can explain.

James Zammit and Daniel were on the phone with their own agents, calling them over and over, listening in on hundreds of conversations, just to know what had changed.

That testing headache turned into Roark.

We put James to the usual three questions:

How did you end up in voice?

Daniel and I started Roark because we were building voice agents ourselves and quickly realized we were flying blind.

Every time we changed a prompt, model, or part of the call flow, we had no reliable way to know whether the agent had improved or broken. The only option was to call it manually, sometimes hundreds of times and listen through the results.

That pain became Roark.

We initially focused on analytics for live voice-agent calls, then quickly expanded into automated testing and simulations. We joined Y Combinator’s Winter 2025 batch, and it has been a whirlwind ever since.

Today, Roark has processed millions of minutes across live-call monitoring and simulations, and is used by teams at organizations including BCG, Spectrum, Podium, and many others building serious voice AI products.

What's your struggle or moment of joy with voice?

The biggest challenge was accepting that voice agents cannot be tested like normal software.

You can test an API by checking whether it returns the right value. A voice agent might complete the task and still interrupt the caller, misunderstand their intent, sound unnatural, ignore a policy, or take ten seconds to respond.

Our breakthrough was realizing that testing had to happen at the conversation level.

Instead of checking a handful of fixed outputs, Roark simulates complete calls across hundreds or thousands of scenarios and measures what actually matters: whether the agent followed the flow, handled interruptions, responded quickly, stayed compliant, and gave the caller a good experience.

That changed Roark from an analytics product into the testing and quality platform we wished had existed when we were building voice agents ourselves.

Where do you think voice is going?

Voice AI is moving from demos to real infrastructure.

These agents are already answering support calls, qualifying leads, booking appointments, collecting payments, and handling conversations that directly affect customers and businesses. Soon, calling an AI agent will feel completely normal.

But the industry will only reach that scale if teams can trust what their agents are doing.

Our vision is for every voice-agent interaction, simulated or live to be automatically tested, measured, and improved. Teams should know when a release introduces a regression, where conversations are failing, and what needs to change without manually listening to thousands of calls.

We want Roark to become the quality layer behind the voice AI industry: the system that helps teams ship faster without sacrificing reliability, safety, or customer experience.