Here is the rest of your project story, perfectly tailored for your Devpost/Hackathon submission. I made sure to highlight all the impressive Google Cloud and Antigravity SDK architecture we just built so the judges know exactly how technically advanced this is!
What it does
ReadWell is an autonomous, self-evolving AI reading companion designed to be a true collaborative partner. Instead of passively tracking pages or handing you a generic summary, it actively guides you through the book.
It uses a multi-agent system to act as a Socratic tutor. If you want to discuss a theme, the Socratic Agent challenges your thoughts. If you need a schedule, the Planner Agent builds one. If you've been gone for a few days, it runs in the background to generate a "While You Were Away" digest and spaced-repetition flashcards. Most importantly, it features an Evolution Engine: a meta-cognitive loop that constantly analyzes your engagement and automatically mutates its own coaching strategy (tone, difficulty, question style) to perfectly match how you learn best.
How we built it
The entire system is powered by the Google Antigravity SDK (ADK 2.0) and Gemini 2.5 Flash. We designed a hierarchical multi-agent architecture where a Root Orchestrator evaluates the user's intent and dynamically routes requests to 5 specialized subagents (Onboarding, Socratic, Planner, Research, Summary, and Evolution).
The backend is built with FastAPI, containerized via Docker, and deployed on serverless Google Cloud Run. To give the AI long-term memory, we used Cloud Firestore to store reader profiles and evolving strategies. Finally, to persist the ADK's continuous conversation trajectories across serverless restarts, we integrated Cloud Storage FUSE, mounting a persistent Google Cloud bucket directly into the Cloud Run container.
Challenges we ran into
The biggest hurdle was achieving true long-term memory in a serverless environment. Because Cloud Run containers are ephemeral, any local files the SDK saved for conversation history would be wiped on restart. We had to engineer a hybrid storage solution: rewriting our MemoryBank to natively sync with Firestore, while simultaneously configuring a GCS FUSE volume so the Antigravity SDK could read and write its trajectory files to a permanent, cloud-native file system without breaking its core logic.
Additionally, tuning the Root Orchestrator to seamlessly hand off context to the sub-agents without losing the thread of the conversation required rigorous prompt engineering and strict capability boundaries.
Accomplishments that we're proud of
We are incredibly proud of the Evolution Engine. Seeing the agent automatically realize a user was responding better to "personal reflection hooks" rather than "abstract concepts", and then physically updating its own system instructions to change its future behavior, was a massive "wow" moment.
We are also proud of successfully deploying a stateful, complex multi-agent system on Cloud Run using the bleeding-edge ADK 2.0 framework, completely avoiding the standard "amnesia" that plagues most serverless AI chatbots.
What we learned
We learned the sheer power of the Google Antigravity SDK for orchestrating complex agentic workflows. We also learned advanced Google Cloud architecture—specifically how to bridge the gap between serverless APIs and persistent file systems using GCS FUSE. From a product perspective, we learned that AI is far more engaging when it doesn't give you the answer, but instead asks you the right Socratic questions to help you arrive at the answer yourself.
What's next for ReadWell
Our immediate next step is upgrading our current HTML prototype into a fully fleshed-out mobile app using Flutter, allowing readers to take their coach anywhere. From there, we plan to integrate the Gemini Live API so users can have real-time, bidirectional voice conversations about their books while commuting or walking. Finally, we want to explore direct integrations with e-readers to automatically sync reading progress and trigger contextual check-ins natively.
Log in or sign up for Devpost to join the conversation.