About

We built the test lab voice agents always needed

Voice agents were being shipped against text tests, log diffs, and one engineer dialing the demo line. Nora and Julian had both watched that fail at scale. Vexa is the instrument they needed and could not find.

Why Vexa exists

Testing voice agents was broken

Not for lack of trying. The tools were built for the wrong medium.

A voice agent fails in ways a text test cannot catch. It talks over the caller. It pauses too long. It handles a clear sentence in a British accent fine, and collapses on the same sentence spoken faster or softer. It understands the words and misses the intent. Every one of those failures sounds obvious in a post-mortem and is invisible in a log file.

For years, the standard approach was to evaluate the language model underneath the agent and ship. If the LLM eval passed, the agent shipped. The problem is that voice is not text. The speech-to-text layer, the response latency, the text-to-speech rendering, the barge-in logic: any one of them can break a call that the model eval never touched.

Nora spent years measuring whether speech recognition worked in production at a large conversational platform, where knowing whether a call actually worked was the hardest unsolved problem. Julian ran the voice and telephony infrastructure behind millions of concurrent calls, and learned that the failures that matter most are the ones no dashboard shows.

They arrived at the same conclusion from different directions: you cannot know whether a voice agent works without running the call. Not a log of the call. The call. Vexa runs thousands of them, across voices, accents, and edge cases, scored turn by turn, before a single real caller hears the agent.

Mission

Every voice, tested. Before a single customer calls.

That is the whole job. Every voice agent deployment should know it holds up across the voices and situations it will actually face, before it faces them.

Most voice agents are tested by the people who built them, on a device they know, in a quiet room, using the exact phrasing they had in mind. That is not a test. It is a rehearsal. The real test is the caller who mumbles, calls from a car, speaks with an accent the team never heard, and does not follow the happy path.

Vexa builds that test. Not as a manual task someone runs before a big release, but as an automated part of every deploy, running the same suite your QA team would run if they had unlimited time and a thousand different voices. The results come back as scores you can read, regressions you can act on, and verdicts you can gate a release on.

What we believe

Four things we hold to

These are not goals. They are the premises the product is built on, and the filter every decision passes through.

Voice failures happen in context

A bad call is not a bad sentence. It is a wrong pause, a missed barge-in, an accent the model never heard. The only way to catch it is to run the call, in full, the way a real caller would experience it.

Simulated beats scripted

A script tells you whether the agent followed a path. A simulated caller tells you whether the agent handled a conversation. These are different questions, and only one of them reflects what happens when a real person calls.

Every provider decision should be reversible

Model-agnostic is not a feature. It is the condition under which a voice team can make honest decisions about their stack without losing their scores, suites, or history when they switch.

Scores that gate deploys are worth building

If a test result cannot stop a broken release, it is documentation, not quality. We build toward verdicts you can act on, including verdicts that say no.

The founders

Built by people who lived the problem

Nora Halvorsen and Julian Reyes are full-time in San Francisco. They founded Vexa in 2024 after spending their careers on either side of the voice quality problem.

Nora Halvorsen

Co-founder and CEO

Speech scientist. Previously led speech recognition evaluation at a large conversational platform, where measuring whether a call actually worked was the hardest unsolved problem. PhD in speech processing.

Julian Reyes

Co-founder and CTO

Real-time infrastructure engineer. Previously ran the voice and telephony platform behind millions of concurrent calls, and learned that the failures that matter are the ones no dashboard shows.

Nine people, full-time, headquartered in San Francisco. Founded 2024. Pre-Seed stage.

The company