Roark: knowing when your voice AI is ready

A voice agent can sound convincing in a demo. The harder test comes when a customer interrupts, changes their mind, or asks to speak to someone. For the person running that operation, the question is whether the call ends with the job done—and how to find out when it does not.

Roark helps teams answer that question. It tests voice agents before release, evaluates real conversations after launch, and helps teams trace failures back to something they can improve. It works alongside an existing voice AI setup, including platforms such as Retell, Vapi and LiveKit. [1]

The company behind the quality checks

Roark was founded in 2025 by James Zammit, CEO, and Daniel Gauci, CTO, and joined Y Combinator’s Winter 2025 batch. Zammit previously worked on infrastructure and AI at AngelList; Gauci’s background includes Akiflow and Casumo. [2]

Their public launch story gives the problem a more personal origin. While building a voice agent for a dental clinic, they encountered callers getting stuck in loops and problems around insurance confirmation. Testing a fix meant calling the agent themselves or working through transcripts. Roark grew out of the need to make that process repeatable. [3]

A quality loop around the agent you already run

Roark brings several related jobs into one platform: prepare the agent for difficult calls, understand what happened in production, and give the team a way to check its own evaluation criteria.

Audio matters here. A transcript may preserve the right words while losing a long silence, a mispronounced name or an interruption. Roark’s post-call analysis evaluates the recording as well as the conversation, then connects findings to issues, alerts and reports. [4]

Three situations where the work becomes concrete

Checking a prompt update before it reaches the phone queue

A team changes a booking agent to handle one difficult request. It still needs to know whether the change affected routine bookings, callers with different accents, or people speaking over the agent. Roark describes simulation suites that can be rerun after prompt or model changes, with configurable callers, background noise and adversarial scenarios. [5]

The value for an operations leader is a repeatable release check. Rather than relying on whichever calls someone remembered to make, the team can define the conversations that must keep working and review the failures before widening a rollout.

Finding support calls that ended without a resolution

A call ending is not always a problem being solved. In its public customer-support examples, Roark focuses on missed handoffs, repeated questions and confident answers that do not match a company’s policy. Those are useful things to examine alongside a dashboard showing how many calls the agent handled. [6]

Our read is that a support team should judge this against its own definition of a successful call: the request was completed, or the customer reached a person with enough context to continue. A high automated score is more useful when a supervisor can listen to the underlying call and understand why it passed.

Reviewing patient-intake conversations against a defined process

Roark’s healthcare page describes checks for identity verification, required notices and pronunciation during patient calls. These are examples of a tightly specified workflow: the team needs to know whether particular steps happened, and what the caller actually heard. They also illustrate why an evaluation rubric needs input from the people responsible for the service. [7]

Roark’s human-review tools let those experts label recordings, resolve disagreements and compare their assessments with automated metrics. For a buyer, that provides a practical way to ask whether the scoring reflects the standards their own team would apply. [8]

Roark’s broader approach links a discovered issue to a simulation, a proposed fix and a check after deployment. The team reviews and applies changes; the platform supplies evidence to help decide whether the work improved. [1]

What its public customer examples tell us

Roark’s website highlights work involving Podium, RadiantGraph and BCG, describing support and sales conversations, healthcare checks, and simulations before deployment. Those examples show the range of work the company is targeting. They do not provide a common before-and-after measure of customer outcomes, so we would use them to guide a demonstration on a buyer’s own calls. [1]

Where Roark stands on SOTA2

Roark ranks first in SOTA2’s Evaluate AI Systems comparison for Technology & Software. The comparison covers 18 candidate products from 11 companies, with six products receiving ranked positions. Its leading position reflects the direct fit between conversation simulations, evaluation and the job of checking whether an agent is ready to release.

Snapshot: October 14, 2026, 10:03 p.m. PT; checked October 15. SOTA2 ranks available evidence using AI review and a weighted rubric for task fit, adoption, product detail, pricing and market fit. This is a task-specific comparison, not a hands-on benchmark of customer outcomes or a ranking of every voice AI product. View the comparison and methodology.

What we would look for in a pilot

We would begin with one call type and a set of recordings that a supervisor has already reviewed. Compare Roark’s findings with that assessment, then test a proposed agent change against the same scenarios. This makes the first question specific: does the tool identify the failures this team actually needs to catch?

For a COO, we would track repeat calls, failed handoffs and the time needed to investigate issues. For a CFO, we would consider the cost of evaluation alongside manual quality review and the work created by unsuccessful calls. Neither needs to accept a single quality score as the business case; both can ask to see the conversations and the operational change behind it.

This spotlight uses public product pages, YC’s public company and launch information, and SOTA2’s ranking snapshot. Product capabilities and customer examples are described by Roark; SOTA2 has not independently audited customer outcomes. The use cases above are documented scenarios, not customer testimonials.