Gloo AI Hackathon 2023 - Kidechisms!

Fostering connection between parents and children by bridging abstract concepts and childlike curiosity through AI-generated personalized stories. $100k Grand Prize Winner, Gloo AI Hackathon 2023.

Teaching foundational values and life concepts is vital for many families, yet the source material is often dense, abstract, and difficult to translate for toddlers or young children. Historical educational texts like the Westminster Larger Catechism contain valuable lessons, but parents frequently lack the tools to facilitate these teaching moments effectively.

Based on insights from families, three critical barriers were identified:

Parents needed confidence that the educational content was grounded in vetted, trusted curriculum rather than AI-generated content that could misrepresent important concepts.

How do you explain abstract ideas like "purpose" or "planning" to a 4-year-old in a way that sticks while maintaining fidelity to the source material?

Traditional formats (reading from a book, memorization drills) feel like work, not connection. Parents needed a way to turn learning into quality time together, to form memories rather than just transfer information.

At the time of the project (October 2023), no solution existed that combined vetted curriculum with AI-generated, personalized storytelling for this context.

A critical constraint was faced with the AI models available during the hackathon: consistency. An inconsistent character disrupts the narrative arc, breaking the child's immersion and turning a moment of connection into a moment of confusion. A design solution was needed that worked the limitations of current generative AI, not against them.

Early explorations leaned into a DJ "remix" metaphor, where parents could spin together curriculum questions and their child's daily experiences to create custom story "tracks.”

However, due to technical constraints, the AI struggled to render the same character consistently from one page to the next. A sheep with a red scarf on page 1 would become a completely different sheep on page 2. Rendering the landscape backgrounds had similar issues.

It became clear that trying to force the AI to create a "perfect" static picture book was a losing battle. Generating a convincing image for every educational scene was too hard to achieve consistently with the available models.

The mental model was shifted from a "Picture Book" (passive consumption) to a "Sticker Book" (active participation).

This design pivot solved two problems simultaneously:

By isolating characters as "stickers" separate from backgrounds, a fine-tuned Stable Diffusion XL model (using LoRA) could be used to generate consistent character assets, while a standard LLM generated atmospheric backgrounds.

The child was shifted from a passive observer to an active participant. Education theory suggests that active learning (dragging and dropping) increases retention compared to passive listening.

Because stickers in real life are rarely placed "perfectly," the AI's inability to render perfect spatial composition became irrelevant. The child's "messy" placement feels authentic and playful. It's a feature, not a bug.

A design system was created to help establish reusable components for developers, ensuring visual consistency and efficiency under time pressure. The components evolved as the product direction evolved, as well.

A visual identity was established that balanced retro playfulness with approachability.

The color palette drew from retro-inspired pastels: soft pinks, warm coral, mint greens, and sky blues. These colors felt inviting rather than overstimulating for young children. They evoke a nostalgic, storybook quality while maintaining enough vibrancy to capture attention and create delight.

Rounded, friendly typefaces were selected that complemented the retro vibe without sacrificing legibility on tablets.

The combination of soft colors and approachable type created a warm, low-pressure learning environment where parents felt comfortable and children felt excited to engage.

Instead of generating a complex scene all at once, a two-stage inference procedure was created:

An LLM parses scene descriptions to generate atmospheric backgrounds that set the visual context.

A fine-tuned version of Stable Diffusion XL using LoRA was utilized to generate base character stickers with consistent visual identity.

A text+image-to-image generator was applied to steer specific expressions (happy, sad, surprised), allowing for emotional nuance without needing the AI to place the character perfectly in the scene.

This approach maintained character consistency while giving the interaction model the flexibility it needed.

A cohesive educational experience was designed where the parent facilitates and the child actively participates. The flow transforms a lesson into a moment of play and wonder.

Parents select a curriculum question and visual theme (e.g., "Sheep in Pasture" or "Outer Space").

The parent reads the AI-generated rhyming story aloud to their child.

The child drags character stickers with different expressions onto scenic backgrounds to complete each page.

At the end, a brief closing statement ties the playful interaction back to the lesson and transitions focus back to the parent-child connection.

The project was recognized for its ability to apply complex AI to a tender, family-centered learning context. The $100k Grand Prize for "Best Overall Value to the Ecosystem" was awarded against 41 other teams at the 2023 Gloo AI Hackathon.

By accepting that perfect scenes couldn't be generated, a more engaging interaction model was discovered. The "Sticker" pivot proved that technical limitations often reveal better UX solutions. Sometimes the constraint the creativity. What initially felt like a failure became the most delightful part of the experience.

The pivot changed the user flow from passive listening to active dragging and dropping. This shift wasn't just about engagement metrics. It was about creating moments where parent and child build something together. The "messiness" of a child's sticker placement became part of the charm, a visual record of their participation and understanding.

The most successful aspect of this project was the recognition that the best learning tools don't replace human connection—they facilitate it. By requiring both parent and child to participate, a ritual of co-creation and wonder was created, not just an app. The digital tool became a catalyst for real moments: the conversation, the laughter, the shared discovery happening between two people.