Most AI product development stories involve a proof of concept that takes months to graduate into something real. CETA Global didn’t have that runway. They had a Google Accelerator demo in December 2025, a vision for an AI product that had never been built before, and roughly two months to get there.
As part of a cross-functional team, AgilityFeat through its WebRTC.ventures real-time communication division, delivered a fully working product tested by real users in three countries before it hit the stage.
The platform is EBT-Sim, a live AI roleplay simulator that lets psychologists practice difficult therapeutic conversations with an AI-generated patient, receive real-time coaching from a second AI agent, and walk away with timestamped, replayable feedback on their clinical performance. CETA calls it their flight simulator because it is the same idea as pilot training: you build competency in high-stakes scenarios before you’re ever responsible for real lives.
That product required product thinking, frontend and backend engineering, cloud architecture, a real-time voice pipeline, and practical multi-agent AI implementation all working together from day one.
The Client and the Problem
CETA Global trains psychologists worldwide in the Common Elements Treatment Approach (CETA), an evidence-based, transdiagnostic mental health treatment protocol designed to work across conditions, countries, and care settings. Demand for CETA has grown significantly across Africa, South America, and other regions. The human-centered training model couldn’t keep up.
The bottleneck was twofold. The number of qualified CETA experts available to train new practitioners was small, and as more organizations sought access to the program, the same small group of specialists had to stretch further. In some cases, countries waited months before training could begin.
The second problem was feedback. Therapy sessions are confidential, which means practitioners had limited ways to get coaching on how they were actually applying CETA once they started working with real patients. Skill development happened slowly, largely without outside input.
How We Engaged
AgilityFeat placed a senior full-stack engineer with deep WebRTC and AI expertise directly into the CETA Global team, working on architecture, clinical alignment, and day-to-day product direction throughout the build.
This is one of the many ways AgilityFeat can work with teams building ambitious AI products: Nearshore Staff Augmentation for embedding specialized engineers directly, and AI Product Development for teams that need a full build from concept to production.
The Build: AI Product Development Across the Full Stack
The approach was to treat EBT-Sim as a complete learning loop, not a chatbot with a clinical skin on it. The goal was to recreate the most important parts of supervised practice in software: a realistic client conversation, live guidance during the session, protocol tracking, structured exercises, and meaningful feedback afterward.
A psychologist enters the platform, selects a chapter and session, starts a live audio or video conversation with an AI client, and works through the CETA protocol without involving a real patient. The system surfaces coaching hints when the trainee drifts off track. If they go too far off course, the session ends early and recommends educational material before they try again. When session goals are met, a separate AI agent evaluates the full interaction and produces detailed feedback with timestamps and replayable moments.
Difficulty was defined by the type of client being simulated, such as resistant, emotionally complex, or avoidant. That allowed trainees to practice not just the steps of CETA, but how to apply those steps when a real patient isn’t making it easy.
AI Architecture
The instinct with LLM products is often to write one large prompt and ask the model to do everything. That breaks down quickly when different parts of the experience have different jobs and different failure modes. This application uses six purpose-built Gemini-powered AI agents through Google ADK, each owning one responsibility:
- simulated client,
- real-time coaching,
- protocol tracking,
- progress assessment,
- whiteboard interaction, and
- post-session evaluation.
Each agent can be tuned and tested independently without one change breaking something else.
Underpinning this is a structured content layer with scenario definitions and protocol configurations that keep AI behavior grounded in clinical standards without asking the models to carry that knowledge on their own.
Tech Stack
- Frontend: Next.js, React, TypeScript, Tailwind CSS, Radix UI, Zustand, WebAudio, Excalidraw
- Backend: Python 3.12, FastAPI, WebSockets, multi-agent session orchestration via Google ADK
- AI / LLMs: Google ADK, Gemini 2.5 Flash, Gemini 2.0 Flash, Gemini 2.5 Flash TTS, Vertex AI
- Speech: Google Cloud Speech-to-Text v2 (Chirp 3), Google Cloud Text-to-Speech
- Infrastructure: Google Cloud Platform, Cloud Run, Cloud Build, Docker, Artifact Registry, Secret Manager, PostgreSQL on GCP
- Observability: Arize / OpenInference tracing
Results
The platform was completed before demo day with the full training loop live: audio and video roleplay, real-time coaching, protocol tracking, whiteboard exercises, and post-session evaluation. Real users in the United States, Chile, and South Africa tested it before launch.
Google responded positively at the demo and later accepted EBT-Sim into the Gemini for Good program. Expansion is now underway across more than six CETA chapters and more than twenty elements per chapter. EBT-Sim is also being integrated into the application CETA practitioners already use to register patients and write session notes, moving from accelerator demo into the core CETA commercial offering.
The Broader Pattern in AI Product Development
EBT-Sim is one example of something AgilityFeat builds regularly: a founder or organization has deep expertise in a domain whether that be a clinical protocol, a training methodology, a compliance framework and needs a team that can turn that expertise into a working AI product, fast.
The challenge is never just the AI, and it’s never just the software. It’s finding a team that can hold both at once: product thinking alongside engineering execution, domain alignment alongside technical architecture.
Ready to Build?
If you have an AI product idea and a team that needs to move fast, we’d like to hear about it.
- Builder Pods — Small, senior, AI-powered nearshore teams for rapid product launches and proofs of concept.
- Full Product Development — Turnkey engagements with our in-house nearshore team, from vision through production.
- Nearshore Staff Augmentation — Add experienced AI engineers to your existing team.
We’d love to hear about the problem you’re solving.
WebRTC.ventures, an AgilityFeat company, also published a technical deep-dive on EBTSim’s real-time communication and AI architecture. Read it here: CETA AI Flight Simulator: AI-Powered Conversation Training and Roleplay Simulation.
