LLMs have changed the paradigm for Voice AI. Compared to rule-based systems (Siri, Alexa, Amazon Polly), LLM-based Voice AI agents understand the intent of the caller and can more often resolve the issue without escalation to a human agent. Moreover, with LLM-based Voice AI agents, developers can build a Voice AI agent more quickly, onboard customers quicker, and iterate on the product faster. Our customers’ Voice AI agents are doing amazing things. It’s so much fun to see the agents achieve the caller’s objective, even in the face of skepticism from the caller.
But LLM-based Voice AI agents are still nascent in some ways. For example, it’s hard for developers to know how their agents are performing. Most Voice AI agent developers are manually listening to calls to identify issues in them. Or they’re finding out about issues with their agent when their customers complain. There’s a better way.
When my cofounder, Adrian Cowham, and I started the company, we were building a semantic cache. We started meeting a lot of Voice AI developers because they were interested in latency improvements from caching. However, we kept hearing them say, “We don’t need to optimize our agent yet. We just need to get it to be more reliable.”
So we pivoted. We’re now building Mixpanel for Voice AI agents. We map caller journeys. We provide audio metrics (i.e., latency) and conversational metrics (i.e., identify calls that end abruptly). We help Voice AI developers improve their agents.
We’d love it if people in the Hacker News community would try out our product and let us know what they think!
Tom
What's your sense on the importance order of the various metrics? I.e when would WPM matter more than SNR or vice versa?
Also, I think it might be worth considering moving the metric explanations up towards the top of the page or adding tool tips. I haven't worked with voice AI agents before and was initially missing some context there.
Thank you for the feedback. We'll move the metrics explanations up!