Roark

Roark is a QA and evaluation platform built specifically for voice AI agents, not another agent builder. It closes the loop between testing and production: simulate an agent against hundreds of callers before launch, score every live call once it is running, and use the findings to ship a proven fix. The team behind it is backed by Y Combinator and reports SOC 2 Type II along with HIPAA BAA support.

Simulation is the pre-launch side. Roark runs your agent against realistic personas built from your own call types, including the angry caller, the rambler and the interrupter, with native accents, code-switching and background noise across 45 languages. Adversarial red-team callers probe for prompt injection and social engineering, while load and health tests confirm the agent survives peak concurrency and stays reachable after launch.

The second half is observability. Every production call is scored as it arrives using Roark's purpose-built audio models rather than a generic LLM grading a transcript, measuring pronunciation, accent clarity, emotion, vocal stress, pace, interruptions and dead air alongside resolution, empathy, hallucination and compliance checks. Issues are filed automatically, dashboards and OTEL traces are available, and changes are diffed so you can see that a fix moved the metric without breaking a caller somewhere else. Roark also publishes Live Bench, a free public leaderboard that scores 300+ speech-to-text, LLM and speech-engine combinations over the real phone network.