Most AI applications operate in turns: a user submits input, the model processes it, and a response is returned. A delay of a second or two goes unnoticed.
Live voice AI does not offer that margin. In a roleplay training session where a psychologist practices a difficult client conversation with an AI-simulated patient, every unnatural pause undermines the realism of the exercise. And with it, the training value.
This was the central engineering challenge behind EBTSim, the AI-powered clinical training simulator that WebRTC.ventures built with CETA Global. CETA Global trains frontline mental health providers worldwide and needed a scalable way for practitioners to build competency in complex clinical conversations before working with real patients. They describe the platform as their “flight simulator”: high-stakes skills, practiced safely, as many times as necessary.
| This topic will be explored in depth Wednesday, July 22 on WebRTC Live Episode 115: Using Voice AI and Avatars to Scale Live Training. |
Six agents, one live session
EBTSim is not a single AI holding a single conversation. It is a multi-agent voice AI system: six specialized Gemini-powered agents, orchestrated through Google ADK, operating concurrently within one live session. One agent plays the simulated client, maintaining a consistent persona and an adaptive trust state that responds to the practitioner’s technique. A second monitors the conversation and surfaces real-time coaching hints. Others track protocol adherence, assess session progress, update a shared whiteboard, and generate a timestamped, replayable feedback report the moment the session ends.
Supporting all of this is a WebSocket-based real-time communication layer that streams audio, transcripts, AI responses, coaching signals, and avatar actions simultaneously — within a latency budget tight enough that the practitioner never perceives the system at work.
From architecture to accelerator
The platform went from concept to production-ready in roughly eight weeks, in time for a December 2025 Google.org Accelerator: Generative AI demonstration. It was tested by users in the United States, Chile, and South Africa prior to launch, and CETA Global has since been accepted into the accelerator program. WebRTC.ventures continues to work with CETA to integrate EBTSim into their broader practitioner ecosystem.
For teams evaluating multi-agent architectures for voice applications, EBTSim offers a clear proof point: the difference between a compelling demo and a tool professionals trust is real-time engineering — disciplined latency budgets, thoughtful agent decomposition, and interaction design that respects how humans actually converse.
