About Roark
We learned the hard way how hard testing is.
So we built Roark.
A fully remote team across San Francisco, London and Malta, building the quality loop for voice AI agents.
Where we come from
Two different industries, one identical scar.
James spent years at AngelList putting large language models to work in financial services — a world where the model reading a decimal point wrong out loud is a regulator’s problem, not a UX bug. Daniel was on the other side of Europe building multi-modality agents for iGaming platforms and then at Akiflow — agents that listened, spoke and acted for people who would never file a bug report, they would just leave.
Different products, different stakes, the same scar: the hardest part was never getting the agent to work in the demo. It was knowing whether it still worked on Tuesday, on a bad phone line, in an accent the test suite had never heard, after the fourth prompt tweak that week.
The hard lesson
Transcripts lie. Demos pass. Production decides.
We both tried to solve it the way everyone tries first: read the transcripts, grade them with an LLM, eyeball a dashboard. And we both watched calls that looked perfect on paper go wrong in the ear — the mispronounced name, the four seconds of dead air, the apology delivered flat enough to lose the customer anyway. Text told us what the agent said. It never told us what the caller heard.
Testing conversations turned out to be a different discipline from testing code. Conversations don’t have fixtures. They have moods, accents, interruptions and background noise — and the only honest test is another conversation.
The bet
So we built the tool we couldn’t buy.
Roark is the quality loop we both wished existed: simulate the calls before launch, score every production conversation on the audio itself, and prove each fix before a customer ever hears it. Purpose-built audio models, not just an LLM grading text — because the failures that cost you customers live in the sound, not the words.
We’re backed by Y Combinator and we work with teams shipping voice AI in healthcare, finance, insurance and beyond — the places where a bad call is more than a bad review.
Founders
Built by the people who got burned.
The two of us above the Golden Gate — San Francisco
James Zammit
Co-founder & CEO
Previously at AngelList, building LLM systems for financial services — where a number read out wrong on a call isn’t a typo, it’s a compliance incident. James shipped the models, then spent his nights building the harnesses to prove they behaved.
Daniel Gauci Mizzi
Co-founder & CTO
Previously built multi-modality agents in iGaming and at Akiflow — systems that had to listen, look and speak at once, for users who never fill in a bug report. Daniel learned that if you can’t replay a conversation, you can’t fix it.
The team
Sarah Chen
Software Engineer · San Francisco
Tom Vella
Software Engineer · Malta
Priya Sharma
Software Engineer · London
A small team of engineers spread across our three home bases — people who'd rather ship the fix than write the postmortem. We're hiring →
Fully remote
Off the clock, we're the people whose houses you can talk to. Home automation, voice assistants in every room, lights that argue back — we use voice everywhere, which is exactly why we care so much when it doesn't work.
Bring a recording.
We’ll score it live.
See your own agent measured on the audio it actually produced — in the demo, in real time. Stop guessing whether your voice AI works.
support@roark.ai · we reply fast