Clascade

Inspiration

Nobody pays attention to a slideshow. We've all sat through one: teacher clicking through bullet points while the class checks out. Meanwhile every "AI makes you a game" demo we'd seen solved the boredom problem by handing control to the AI and hoping for the best, one unreviewed generation, no fact-checking, no way for a teacher to actually trust it in front of 30 kids. That's the gap. Teachers don't need a toy. They need something they can check, approve, and actually run.

What it does

Clascade takes a teacher's slideshow, or just a chat description of a lesson, and turns it into an interactive 3D lesson the whole class plays together in the browser. Before anything gets rendered, the AI produces a structured lesson spec, not code: every scene, every fact, every camera move lives in that spec first. Facts get checked against real sources with citations the teacher can approve. Content gets screened for age-appropriateness, with any adjustment shown transparently. Then the teacher publishes and runs it live, advancing every student's screen at once, pausing the whole class, pulling up any one student's view, so the class moves together instead of forty kids racing off in forty directions.

How we built it

Spec-first, always. Slides get ingested and broken into concept beats. Each beat gets mapped to one of our hand-built pedagogy templates (cinematic timeline, scale journey, parameter sandbox), and the whole thing gets compiled into a validated JSON lesson spec before a single pixel renders. Fact-checking runs through search-grounded generation with citations attached to every claim. Safety runs through a dedicated review pass. The renderer is Three.js with real physics, built to actually run on a school Chromebook, not just a dev laptop. Classroom sync runs on Firestore: the teacher writes the current phase, every student's screen reacts live. Narration runs in the teacher's own voice from a 10 second sample. The AI never touches rendering code directly; it only ever fills in the spec.

Challenges we ran into

Keeping the AI on a leash was the whole game. It's tempting to just let the model generate a Three.js scene straight from a prompt. It's faster and it demos fine once. But that's a black box no teacher should trust with their class, and it falls apart the second you need to review or edit anything. Forcing every generation through a validated spec instead of raw code made everything slower to build and much harder to get wrong.

Accomplishments that we're proud of

Getting a teacher-controlled classroom to actually work: every student's screen locked to the same phase, nobody able to wander off ahead, a teacher able to freeze the whole room with one click, while the content underneath is still AI-generated and grounded in real sources. That combination, control plus verification, is the part we think nobody else at this hackathon has.

What we learned

Letting the AI write the game directly is way easier than making it prove its work. We could've shipped something flashier a lot faster by skipping the spec layer entirely. We didn't, because a teacher who can't review what their class is about to see isn't going to use it twice.

What's next for Clascade

Real classroom telemetry, tracking which concepts students actually struggled with instead of mocked data. More pedagogy templates beyond history and biology. And eventually, live interaction between teacher and students inside the scene itself, not just phase control from the sidelines.