Simulate realistic callers at scale with custom personas, scenarios, interruptions, noise, accents, and network conditions. NovaSynth runs those calls against your voice agent, scores audio and transcripts across 30+ dimensions, and surfaces the failures and fixes that matter to your team.
My name is Shashank, co-founder and CEO at Noveum AI.
Aditi's comment covers the caller. This one covers the rest of the loop: how a call gets structured, how it gets judged, and what happens to what you find.
A real call does not run top to bottom
A test script is a straight line. A real call is not. A scenario in NovaSynth is a tree instead of a sequence.
The caller asks for a refund. If the agent asks for an order number, the caller gives a wrong one. If the agent offers a callback instead, the caller refuses and asks for a manager. Each branch carries what the agent was supposed to do at that point, so a failure lands on a turn and not on the whole call.
A call can sound perfect and still break a rule
Relevance, tone, and holding role under pressure are matters of degree, so we score them. Whether the agent asked all three qualifying questions before transferring is not a degree. It happened on that call or it did not.
An LLM judge rating the whole call will sometimes note the missed step in its own reasoning and still give the call a high score, because everything else went well. So hard rules get checked as rules on our reports, and judgment gets scored.
Fixing the interrupter without slowing down every other call
Catching the failure is the start of the work, not the end of it. NovaPilot, our fix engine, finds the pattern behind the failing calls, tests 136+ prompt variations across four evolutionary generations, and returns the one that holds as a pull request in your repo.
Every fix is then backtested against your existing calls and traces. Teaching an agent to handle an interrupter can add a beat of latency everywhere else, and in a voice call latency is its own kind of failure.
Happy to go deep on any of it
Aditi and I are both here all day. If you build voice agents, I would like to hear how you test them today. Ask about the audio stack, the scenario engine, or the parts of the product you think we have got wrong.
One question back to you: what is the failure you have only ever caught in production, never in testing?
Report
I am building a voice agent for a niche case, and NovaSynth is exactly what I need to pressure-test my solution before it hits production and starts serving my users.
So excited to finally see this live! ❤️ Team has spent so much time thinking about all the ways a “real” caller can break a voice agent, and it’s amazing to see that turn into something people can actually use. The idea of testing voice agents against all the messy, unpredictable things real callers do is something that makes so much sense once you see it.
Can’t wait to see what kind of crazy caller scenarios people throw at NovaSynth 😄
honestly, this is probably one of the most interesting things i’ve worked on.
novasynth is one of those products where the idea just clicks. you can test your agent with all kinds of conversations, weird edge cases, unexpected turns, and not just the happy path, before real users run into them.
really glad i got to be a part of building this and see how it all came together.
A great way to ensure in prod our voice agents give comparable quality as we got in our pilots !!!
Quash
Pretty great usecase actually. we've also seen these requirements and need for a specific tool among our clientele.
NovaSynth by Noveum
@pr_khar do get in touch to discuss the use cases. We run evaluation for pre-prod as well as voice agents running in production.
NovaSynth by Noveum
Hi Product Hunt 👋
My name is Shashank, co-founder and CEO at Noveum AI.
Aditi's comment covers the caller. This one covers the rest of the loop: how a call gets structured, how it gets judged, and what happens to what you find.
A real call does not run top to bottom
A test script is a straight line. A real call is not. A scenario in NovaSynth is a tree instead of a sequence.
The caller asks for a refund. If the agent asks for an order number, the caller gives a wrong one. If the agent offers a callback instead, the caller refuses and asks for a manager. Each branch carries what the agent was supposed to do at that point, so a failure lands on a turn and not on the whole call.
A call can sound perfect and still break a rule
Relevance, tone, and holding role under pressure are matters of degree, so we score them. Whether the agent asked all three qualifying questions before transferring is not a degree. It happened on that call or it did not.
An LLM judge rating the whole call will sometimes note the missed step in its own reasoning and still give the call a high score, because everything else went well. So hard rules get checked as rules on our reports, and judgment gets scored.
Fixing the interrupter without slowing down every other call
Catching the failure is the start of the work, not the end of it. NovaPilot, our fix engine, finds the pattern behind the failing calls, tests 136+ prompt variations across four evolutionary generations, and returns the one that holds as a pull request in your repo.
Every fix is then backtested against your existing calls and traces. Teaching an agent to handle an interrupter can add a beat of latency everywhere else, and in a voice call latency is its own kind of failure.
Happy to go deep on any of it
Aditi and I are both here all day. If you build voice agents, I would like to hear how you test them today. Ask about the audio stack, the scenario engine, or the parts of the product you think we have got wrong.
Start a Free Trial: https://bit.ly/4zOcEXT
Book a Demo: https://bit.ly/4gtBXae
One question back to you: what is the failure you have only ever caught in production, never in testing?
I am building a voice agent for a niche case, and NovaSynth is exactly what I need to pressure-test my solution before it hits production and starts serving my users.
NovaSynth by Noveum
NovaSynth by Noveum
So excited to finally see this live! ❤️
Team has spent so much time thinking about all the ways a “real” caller can break a voice agent, and it’s amazing to see that turn into something people can actually use.
The idea of testing voice agents against all the messy, unpredictable things real callers do is something that makes so much sense once you see it.
Can’t wait to see what kind of crazy caller scenarios people throw at NovaSynth 😄
NovaSynth by Noveum
honestly, this is probably one of the most interesting things i’ve worked on.
novasynth is one of those products where the idea just clicks. you can test your agent with all kinds of conversations, weird edge cases, unexpected turns, and not just the happy path, before real users run into them.
really glad i got to be a part of building this and see how it all came together.
excited to see novasynth on product hunt today :)
LaunchPedia
This looks great. Very helpful for businesses looking to setup voice agents
congratulations on the launch @harkirat_singh3777 and team.
NovaSynth by Noveum
@karthik_tatikonda Thank you for the support.