Roark Is Building the Quality Control Room for Voice AI
Voice agents now answer phones in clinics, banks, and call centers. Roark, a Y Combinator W25 company, is the tool that checks whether they are actually doing their job - by listening to the calls the way a customer would.
A patient calls a dental clinic to move an appointment. The voice on the other end is not a receptionist - it is an AI agent, and it handles the reschedule cleanly. Then the patient adds, almost as an afterthought, that the tooth still hurts. The agent, trained to book and confirm, sails right past it. The words on the transcript look fine. The call was not.
That gap - between what a voice agent says and what a caller actually experiences - is the whole reason Roark exists. The San Francisco company, part of Y Combinator's Winter 2025 batch, builds observability and testing tools for the AI agents now picking up phones in clinics, banks, insurance lines, and customer support queues. Its pitch is short: test, monitor, and improve your voice agents. The work behind that sentence is where things get interesting.
The problem hiding in plain hearing
Building a voice agent has gotten dramatically easier. Platforms like VAPI and Retell let a small team stand one up in days. Making that agent reliable is the hard part, and it is where most teams are flying blind. The common QA method is embarrassingly manual: an engineer calls the agent a few times, listens, decides it sounds okay, and ships. Real callers then arrive with accents, interruptions, background noise, and questions no one scripted.
Roark's argument is that voice needs its own instruments because it fails in ways text tooling cannot see. A transcript will tell you the agent said the right words. It will not tell you the caller sighed, paused for three seconds, or repeated the question because the answer landed wrong.
So Roark went audio-native. Its evaluation engine scores what callers actually hear, listening for vocal cues - frustration, hesitation, the pauses that signal confusion - rather than grading a clean text log. A single call can be analyzed with up to 15 speakers automatically identified, which matters for the messy, multi-party reality of a real phone line.
What Roark actually does
The quality loopThe platform is best understood as a loop that runs around a voice agent, from staging to production and back into the next fix.
The replay feature is the one engineers tend to fixate on. Instead of re-testing an agent by hand after every prompt tweak, a team can re-run its real conversations against the updated logic and see what broke. It is the voice equivalent of a regression test suite, and until recently it barely existed.
On top of that sit monitoring dashboards, real-time alerts for compliance violations or sentiment spikes, and automated test generation that turns a failed call into a case the agent has to pass next time. The result is less a single feature than a habit: treat voice agents like production software, with the safety nets that implies.
The simulation side deserves its own note, because it inverts the usual order of testing. Rather than wait for a real caller to expose a weakness, Roark spins up synthetic personas - different accents, languages, speech patterns, and moods - and runs them at the agent over phone or WebSocket connections, for both inbound and outbound flows. The cheapest failure, the reasoning goes, is the one you cause on purpose, in staging, before a customer ever dials in. A team can watch how its agent handles an impatient caller, a heavy accent, or someone who interrupts mid-sentence, and fix the behavior before it ships.
Transcript-grading vs. hearing the call
The clearest way to see Roark's angle is to line it up against the default approach most teams start with.
Text-based monitoring
- Reads the transcript, not the audio
- Misses sighs, pauses, tone, and frustration
- Manual re-testing after every change
- Failures found after customers hit them
Roark's audio-native loop
- Scores what the caller actually hears
- Detects vocal cues and sentiment shifts
- Replays real calls against new versions
- Turns failures into automated tests
Who is behind it
Roark was founded in 2025 by James Zammit (CEO) and Daniel Gauci (CTO), both engineers with roughly a decade of experience each. Before Roark, Zammit spent years in infrastructure at AngelList, where he worked on the systems that scaled the platform's portfolio assets from $10B to $124B and helped build an AI-powered portfolio manager. Gauci's background runs through multi-modal agents and earlier startup work.
That resume matters for a company selling reliability. Zammit's earlier career was spent making financial infrastructure trustworthy at scale - the same instinct, pointed at a newer and noisier surface. Companies that sell confidence tend to be founded by people who have already been on the hook when a system failed quietly, and that is the temperament Roark is built on.
The team is deliberately small and spread out - three people across San Francisco, London, and Malta - which shapes the product as much as the roadmap. A tight team building infrastructure tends to obsess over the primitives: the replay engine, the scoring models, the integrations that have to hold up under millions of minutes of real traffic. There is little room for feature sprawl when three engineers are responsible for a system other companies rely on to catch their own mistakes.
Where it fits in the market
Zoom out and Roark is making a familiar bet: every time the software interface changes, the observability stack gets rebuilt. Web apps got Datadog and Sentry. LLM apps spawned a wave of eval and tracing tools. Voice, with its own failure modes - latency, interruptions, dropped context, sentiment - is the next surface to need its own instruments.
The company is not alone in noticing. A small cluster of voice-AI evaluation startups has appeared, and general LLM-observability tools are stretching toward voice. Roark's wager is that audio-native scoring and real-call replay are hard enough to build well that they become the reason teams pick it over a script they wrote themselves.
There is a defensibility argument buried in that wager. Anyone can write a script that dials an agent and checks the transcript. Far fewer can build models that reliably hear frustration, separate 15 speakers on a noisy line, and re-run thousands of real conversations against a new agent version without a human in the loop. The moat, if there is one, is in the audio - the part that does not show up in a text log and is genuinely hard to fake.
The business, and the traction
Roark sells a B2B SaaS subscription to teams building voice agents, with value scaling to the volume of calls monitored and simulated. Its early customer base leans on the network around it: fellow Y Combinator startups, plus companies in healthcare, finance, insurance, legal, home services, hospitality, and support. The reported headline number - more than 10 million minutes of calls processed in the first six months - is the kind of figure infrastructure companies accumulate quietly until the volume gives them away.
Those industries are not a random spread. They are the places where a botched call carries real cost: a missed symptom in healthcare, a compliance slip in finance or insurance, a lost booking in home services. For a clinic or a bank, a voice agent that sounds fluent but mishandles one call in twenty is not a convenience, it is a liability. That is the buyer Roark is built for - the team that has already put an agent on a live phone line and now has to answer for what it does there.
The team is still small, three people spread across San Francisco, London, and Malta, and backed by Y Combinator's W25 batch. Funding details reported publicly are modest and early-stage, which fits a company at this phase: the interesting metric is call volume, not the raise.
Whether Roark becomes the default QA layer for voice AI or one option among several will come down to the usual things - how deep the integrations go, how much the replay-and-improve loop saves teams in real incidents, and whether voice agents keep spreading fast enough to make dedicated tooling non-negotiable. For now, the bet is straightforward: as more of the phone calls in your life get answered by software, someone has to be listening to make sure that software is any good.