We built RubricHQ because we lived the problem.

The problem we couldn't ignore

We were building voice AI for customer support — the kind that handles real phone calls, real customers, real stakes. And every time we shipped a prompt change, the same thing happened: we'd test 10 calls manually, it would sound fine, and then we'd deploy.

Three days later, someone would Slack us: "The agent is promising 30-day refunds again." Or: "Latency spiked to 3 seconds and completion rates dropped 40%." Or the worst one: "A customer recorded the call where our agent gave medical advice."

We'd scramble to reproduce the issue, roll back the prompt, and promise ourselves we'd test more carefully next time. But "more carefully" meant calling our own agent 20 times instead of 10. It was never enough.

The realization

Software engineers don't ship code without tests. They don't deploy without CI. They don't merge without code review. But voice AI teams? We were deploying conversational agents that handle thousands of calls a day — with no automated testing, no regression detection, and no way to measure quality beyond "it sounded fine to me."

That gap is what RubricHQ exists to close.

What we're building

RubricHQ is the testing and evaluation platform for voice AI. We simulate thousands of voice calls against your agent, evaluate every transcript with LLM-as-judge and audio metrics, and give you the confidence to ship prompt changes without praying.

We work with teams building on Vapi, Retell, LiveKit, and Pipecat. Whether you're running 50 calls a day or 50,000 — if your agent talks to real people, it needs real testing.

Our belief

Voice AI is going to handle most of the world's phone calls within a decade. The teams building these agents deserve the same quality tooling that web developers have had for years. Automated testing, regression detection, metrics dashboards, version control for prompts — none of this should be optional when your agent is talking to real humans.