A few weeks ago we came across the Gemini Live Agent Challenge, a Google hackathon focused on building real-time AI agents with Gemini. We were immediately interested — we had been experimenting with text-based AI apps for a while, but we had always wanted to build something with live agents.
At first we explored a few ideas around language learning, but eventually we landed on something that felt much more personal: Rehearse AI.
Rehearse AI is an app designed to help people practice difficult conversations. You can simulate a conversation with an AI agent, rewind the timeline, try different responses, and explore how the conversation could unfold depending on what you say.
The idea is simple: practice the conversation before it happens in real life — so when the moment comes, you’re ready.
We started prototyping the project in Google AI Studio and built the app using Gemini models, the Google AI SDK, and React.
One of the biggest challenges was designing the conversation timeline. Because users can rewind and try different responses, we had to figure out how to reflect those changes both in the UI and in the AI’s conversational state. That meant iterating on how the timeline works, how branches are represented, and how the AI understands the current context of the conversation.
Over the past three weeks we’ve been building, testing, and iterating on this idea.
What excites us most is the potential impact. Many people experience anxiety when facing difficult conversations — whether it’s asking for a raise, giving feedback, or setting boundaries. Rehearse AI creates a safe space to practice those moments.
For us, it’s also been a new way to think about AI: not just as a tool to generate text, but as a simulation environment for real-life decisions.
This content was created as part of our participation in the Gemini Live Agent Challenge hackathon.
Top comments (0)