Launch HN: Vocera (YC F24) – Testing and Observability for Voice AI

I've been working with auto-generated content for the past 8 years (both algorithmic and LLM-based). One of the biggest challenges is detecting and preventing regressions after "improving" the prompt, algorithm, or model.

Just yesterday, I deployed a voice agent (OpenAI + Twilio), and it's clear that there are countless scenarios where the agent might not perform as expected. For example, if you ask the agent to speak in German, but your tool uses call names or returns data in English, the agent might suddenly switch to speaking English.

Overall, I believe voice agents will primarily be used and developed by SMEs, but they often lack the time or expertise to account for all these edge cases.

Btw: here is the number agent. sorry it's in German: +43732 350011

One trend I’ve noticed is there’s a really heavy focus on pre-deploy tests which makes a lot of sense. But one big gap is the lack of ability to surface the scenarios that you don’t know are even occurring after deployment. These are the scenarios that human agents are great at handling and ai agents often fall flat: in a sales context that can have a direct impact on a customers bottom line.

I think rather than attempting to deploy a perfect agent, having a mechanism to surface issues would lend much more peace of mind when launching ai voice agents. Would be happy to chat more if additional context/real world examples would be helpful. Congratulations again on the launch!

Background: have worked on contact center unstructured analytics for several years.

Agree to everything you said. That is why we have our observability platform, which allows your live calls to be monitored. The idea is to use the observability platform to run real-life simulations so as you make fixes, you can test it in simulation environment