Disclaimer: This piece of content was created for the purpose of entering the #GeminiLiveAgentChallenge hackathon.
Sales is a high-stakes performance, but until now, practicing for it has felt like a chore. Most representatives walk into critical discovery calls with high anxiety because they haven't "repped" the scenario in a realistic environment.
Enter Reptrainer (DealPilot): a real-time AI training ground built on the cutting edge of the Google Cloud AI ecosystem.
The Vision: A Safe Space for High-Pressure Conversations
Our goal was to create a "flight simulator" for sales. We wanted to move beyond the robotic, turn-taking chatbots of the past and build something that felt human. The release of the Gemini Live API was the missing link, finally allowing us to build a voice-first agent that is responsive, reactive, and capable of natural interruption.
The Tech Stack: Truly Multimodal
Building a system that listens, thinks, and speaks required a deeply integrated multimodal architecture that minimizes latency and maximizes realism.
1. The Voice Brain: Gemini Live API & Python ADK
The core interaction is powered by a Python backend using the Google Agent Development Kit (ADK) and FastAPI.
- Bidirectional Streaming: We leverage WebSockets to stream audio directly to our ADK-powered agent.
- Realistic Interruption (Barge-In): One of our proudest achievements was tuning Voice Activity Detection (VAD) to ensure the AI "stops" speaking instantly when the user speaks - just like a real prospect would.
2. The Unified Multimodal Debrief
Sales training is only as good as the feedback. Traditionally, generating a visual and narrated report would require stitching together multiple services. Reptrainer achieves this in a single "leap" using Gemini's native multimodal output:
- Unified Generation: Instead of separate calls for text and images, we use a single API call to generate both coaching insights and contextually coherent infographics interleaved in the response.
- Contextual Coherence: Because the model sees the coaching text and the image generation as part of the same task, the visuals directly and accurately reflect the specific feedback given to the rep.
3. Grounded Intelligence with Vertex AI
A "generic" buyer persona is useless for enterprise sales. We use Vertex AI grounding to ensure our personas are aware of your actual product documentation and competitor "battle cards." If you mention a specific feature or rival, the AI knows exactly how to push back based on your real-world market position.
What We Learned (The Hard Way)
Building with Gemini Live taught us several key lessons about the future of voice AI:
- Brevity is King: In voice interactions, long responses feel like monologues. We learned to implement strict brevity guards to keep the flow conversational.
- Managing "Dead Air": Executing tools (like real-time data lookups) takes time. We coached our AI to use natural filler phrases like "Hang on, let me check that pricing for you..." to maintain immersion while the tools execute in the background.
Infrastructure: Scalability with Google Cloud
The entire system is hosted on Google Cloud Run, orchestrated by Cloud Build for a seamless CI/CD pipeline. We use Firebase for real-time state synchronization, Firestore for session persistence, and Cloud Storage for hosting multimodal assets.
The unified Google Cloud ecosystem allowed us to build features in days that would have normally taken weeks of "glue code" with disparate providers.
What’s Next?
We’re just scratching the surface. Our roadmap includes integrating Gemini Vision to provide real-time coaching on body language, and simulating "Buying Committees" where you pitch to multiple AI stakeholders at once.
Building Reptrainer has shown us that the future of sales training isn't just AI—it's Live AI.
Created for the #GeminiLiveAgentChallenge hackathon.
Top comments (0)