Transform any book into a children's picture book.
Let a 6-year-old experience The Great Gatsby — not through a boring, stripped-down summary, but through a genuine picture book that keeps the original soul, characters, and vibe intact.
Google Cloud Rapid Agent Hackathon · Track: MongoDB · Devpost
🔗 Live demo: https://picture-book-gen-e3mtc46uua-uc.a.run.app
When I was a kid, I hated reading long blocks of text and always preferred pictures. Classic literature sounded so boring to me, and history books were just way too long. I wish there were classic storybooks made specifically for young kids — not through a boring, stripped-down summary, but through a genuine picture book that keeps the original soul, characters, and vibe intact.
To do this, we had to solve three big problems:
- Smart Simplification — Breaking down a massive novel into distinct scenes without losing the core plot, and translating it into 6-year-old friendly language.
- Text-to-Image Alignment — Making sure the illustrations actually match the emotions and details of the story.
- Character Consistency — Keeping the main character looking exactly the same across a 40-page book without their face or clothes shifting.
I use Google ADK's SequentialAgent to run four distinct roles back-to-back on Cloud Run:
| Agent | Role |
|---|---|
| Analyzer | Pulls the book in (e.g. from Project Gutenberg), maps out who the main characters are, and uses the TextTiling algorithm to find the story's emotional highs and turning points. |
| Writer | Takes those scenes and rewrites them into simple, engaging prose for kids. |
| Artist | Uses Gemini 3 Flash Image to create the illustrations and blends the story text naturally into the art (like inside clouds or speech bubbles). |
| Vision QA | Powered by Gemini 3.5 Flash (Vision), this agent checks the images against the original character design. If a page doesn't look right, it triggers a self-correction loop to automatically redraw it based on feedback. |
To stop characters from changing appearance, I used the MongoDB MCP Server:
- During the setup phase, every character gets a locked-in "Visual Identity Sheet" saved in MongoDB as the single source of truth.
- Every time the Artist Agent draws a new page, this reference sheet is read back through MCP, ensuring the protagonist looks identical from page 1 to page 40.
- AI Models — Everything runs on Vertex AI: Gemini 3.5 Flash for fast text processing and Vision QA, and Gemini 3 Flash Image for drawing.
- Compute & Storage — The backend runs on Cloud Run, and all image assets, character data, and final PDFs are stored in Google Cloud Storage.
The final result is a beautiful, square-format PDF book, plus an interactive web app where users can click and fine-tune any character, scene, or text overlay on the fly.
- Chaining mini-agents is way better than one giant prompt — it made the whole storytelling pipeline incredibly stable and fast.
- MCP is a total game-changer for database sync, allowing me to lock in a single "visual identity source of truth" without writing endless glue code.
- Closing the loop with automated Vision QA is the future — it turns unpredictable AI drawings into a reliable, high-quality production line.
pip install -r requirements.txt
cd frontend && npm install && cd ..
cp .env.example .env
# Default backend is Vertex AI (uses your gcloud ADC + GCP_PROJECT).
# No GCP project? Set GEMINI_BACKEND=api_key and add GEMINI_API_KEY.
python -m uvicorn src.app:app --port 8000 # backend
cd frontend && npm run dev # frontend → http://localhost:3000This project is open source under the GNU AGPL-3.0.
You are free to use, study, modify, and redistribute it, including running it as a network service — provided derivative works and hosted modifications are also published under AGPL-3.0. For commercial licensing outside the AGPL terms, contact the author.
Built for the Google Cloud Rapid Agent Hackathon 2026 (MongoDB track) with Google ADK · Gemini 3 on Vertex AI · MongoDB MCP.