Inspiration

It started with a memory: a kid stuck on ¾ + ⅝, lost in rules with no meaning until someone drew two pizzas cut into slices, and the fraction suddenly made sense. That one drawing did more than pages of explanation ever did. EduMotion AI exists to be that pizza drawing for any topic, for anyone who learns better by seeing than by reading.

What it does

EduMotion AI turns any topic or question you type into a short, narrated, lesson with a quiz at the end a local LLM writes the script, matching visuals and voice narration are generated, and the whole thing is automatically into a finished video in under a minute, no manual editing required.

How we built it

A local LLM (Qwen3 8B via Ollama) turns a topic into a structured script (title + scenes). Each scene gets a matching icon/illustration, narration via TTS, and is rendered as an animated video using Remotion with spring-based motion and staggered text instead of static slides. A quiz is generated alongside the script for reinforcement.

Challenges we ran into

First attempts looked like a slideshow, not animation fixed with springs, staggered timing, and continuous micro-motion. Getting Ollama to reliably output clean, parseable JSON took real prompt tuning. Small dependency/versioning issues meant testing every layer end-to-end instead of assuming it worked.

Accomplishments that we're proud of

A fully automated pipeline: question in, narrated video + quiz out, no manual editing. Viddeos that feels intentional and TED-Ed-like, not a fade-through slideshow. Reusable scene templates that scale to any topic consistently.

What we learned

Good explanation is a design problem, not just a content problem. Orchestrating multiple AI models around one shared data format is harder and more valuable than any single model. "It renders" and "it looks good" are very different bars.