NovaSynth by Noveum simulates real SIP calls, LiveKit audio, and chat interactions to stress-test conversational AI agents against interruptions, edge cases, and adversarial behavior.
What the product does and how it is positioned
NovaSynth by Noveum is an AI agent simulation platform designed to evaluate voice and chat agents before deployment. By dispatching synthetic users over live SIP telephony connections, LiveKit audio streams, and HTTP chat interfaces, it surfaces behavioral edge cases, latency problems, and turn-taking failures in simulated rehearsal sessions.
The system generates multi-turn scenarios incorporating varied accents, conversational interruptions, background noise, and prompt-injection probes. Each session is recorded as an end-to-end trace and assessed against voice-specific evaluation metrics, while tool sandboxing is designed to support synthetic runs do not commit unauthorized writes to live production systems.
Source-supported ways to use the product
Simulating high-interruption phone calls over SIP connections to assess agent barge-in response and latency under realistic audio conditions.
Sending calculated injection attempts via chat endpoints to verify agent guardrails, system prompt secrecy, and refusal mechanisms.
Scaling call volume up to 2,000 concurrent lines across combinatorial scenario matrices to measure agent infrastructure stability.
The documented workflow, where available
Configure caller personality attributes, background details, language preferences, moods, and interruption tendencies.
Generate or author interaction scenarios and edge cases based on agent context, documentation, and prompt configurations.
Select voice and conversational evaluation metrics to monitor throughout the multi-turn synthetic conversation.
Capture the full call trace, evaluate failure signals, and analyze root causes surfaced by the simulation.
Validate adjustments against the failing scenarios end-to-end to confirm behavioral remediation before production deployment.
NovaSynth avoids testing against mocked transcripts by connecting directly to voice agents over live communication channels. Calls are routed via SIP through providers such as Twilio, Plivo, and Vobiz, alongside real-time audio streams powered by LiveKit and Pipecat. This setup forces synthetic interactions to reflect real-world network latency, background interference, and audio anomalies.
The underlying orchestration infrastructure uses BullMQ, PostgreSQL, and Redis to parallelize runs across distributed queues. Tool interactions within simulated runs leverage virtualization to replay read operations and sandbox write operations, helping reduce synthetic interactions from modifying real production environments.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
NovaSynth can simulate phone calls over SIP via Twilio, Plivo, or Vobiz, real-time audio through LiveKit and Pipecat, and chat over HTTP.
The platform utilizes tool virtualization that replays read actions and sandboxes write actions so synthetic runs do not affect production systems.
NovaSynth supports load testing up to 2,000 concurrent lines, though teams must notify the provider in advance so lines can be provisioned.
Synthetic sessions are captured as traces and scored using over 30 voice AI metrics, drawing from a broader platform library of more than 100 evaluators.
Noveum products can be deployed as SaaS, hosted within a user VPC with Bring Your Own ClickHouse, or installed completely on-premise.
