CloudTutor

Landing page

Overview

CloudTutor is a live AI tutor for Google Cloud. We built it to solve a simple but painful problem: cloud documentation is powerful, but learning from it alone can feel static, overwhelming, and slow. Beginners often do not just need answers, they need a teacher.

Our idea was to create a voice-first tutor that feels more like a real instructor than a chatbot. A user can ask a question such as “What is Firebase Cloud Functions?”, get a concise explanation, interrupt naturally with follow-up questions, and then move into a guided documentation walkthrough. Instead of leaving the user alone with a docs page, CloudTutor narrates the journey, opens the relevant official documentation, and turns self-serve learning into an interactive session.

What inspired us

We were inspired by the gap between powerful cloud tools and the learning experience around them. Many people want to build with Google Cloud, Firebase, and adjacent services, but they struggle to connect product concepts, official documentation, and practical use cases. We wanted to make that process feel more human, more interactive, and more confidence-building.

How we built it

We built CloudTutor as a multimodal system with a Next.js frontend and a Python backend. The core live tutoring experience is powered by the Gemini Live API through ADK, which lets the user speak naturally, receive spoken answers back, and interrupt the agent in real time.

On top of that, we added a guided documentation mode. When the user wants to go deeper, CloudTutor launches a browser walkthrough, opens official Google Cloud or Firebase documentation, and presents a split-screen learning experience. The browser side handles the live documentation session, while the companion panel shows timeline events, navigation state, and walkthrough context.

We also built tutorial artifact generation so a session can produce reusable notes and diagrams after the lesson, turning a live conversation into something the learner can keep.

Challenges we faced

The biggest challenge was making the experience feel truly live instead of turn-based. Real-time interruption, voice playback timing, barge-in behavior, and maintaining context across tutoring + browser navigation required a lot of careful iteration. Another challenge was keeping documentation navigation deterministic so the agent would open the correct official docs for the service currently being discussed rather than drifting into generic search behavior.

We also had to balance autonomy with safety. The browser workflow needed to feel smooth and helpful while still making it clear when the user is approving an action.

What we learned

We learned that building a great live agent is not only about model quality. It is also about orchestration, timing, state management, and user trust. The details around interruption, context handoff, narration, and predictable grounding have a huge effect on whether the experience feels magical or frustrating.

We also learned that multimodal tutoring becomes much more compelling when the agent can combine voice, live reasoning, grounded sources, and visual walkthroughs in one continuous experience.

Why this matters

CloudTutor rethinks cloud education as an immersive, voice-first learning experience. Instead of forcing users to choose between “chat with AI” and “read the docs,” it combines both into one guided flow. The result is a more approachable way to learn complex cloud services, especially for beginners who benefit from explanation, pacing, and interaction.