StorySprout

Classic novels are too dense for kids. Story Sprout fixes that by turning any masterpiece into a gorgeous children's picture book. No language barriers, and no character face-drifting.

Comment

Inspiration

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.

What it does

It takes dense, complex, and unreadable masterpiece novels (like The Great Gatsby ) and automatically transforms them into gorgeous, context-aware children's picture books.

How we built it

The 4-Agent Pipeline

I use Google ADK’s SequentialAgent to run four distinct roles back-to-back on Cloud Run:

  • Analyzer: Pulls the book 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 until it's perfect.

Solving Consistency (MongoDB MCP)

To stop characters from changing appearance, I used the new 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, MCP injects this reference sheet as a tool, ensuring the protagonist looks identical from page 1 to page 40.

Google Cloud Infrastructure

  • AI Models: Everything runs on Vertex AI. We use 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 (GCS).

The Final Result: 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.

Challenges I ran into

To do this, I solved 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.

Accomplishments that I am proud of

  • Anyone can produce high-quality, visually consistent stories without a professional design background.
  • It slashes the time and budget needed for animation and storyboards, opening doors for anyone with a great idea.

What I learned

  • 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.

What's next for StorySprout

I want to build a shared repository where parents and educators can publish their generated children's books, remix other creators' character sheets, and co-create multi-chapter fantasy universes using a unified, crowdsourced visual identity hub.

Built With

  • cloud-run
  • css3
  • gemini-3-flash-image
  • gemini-3.5-flash
  • google-agent-development-kid(adk)
  • google-cloud-storage(gcs)
  • html5
  • javascripts
  • model-context-pprotocol(mcp)-sdk
  • mongodb
  • mongodb-altas
  • mongodb-mcp-server
  • python
  • texttiling-algorithm
  • typescript
Share this project:

Updates