Skip to Chalk showcase

Chalk showcase

A complete generated lesson, ready to explore.

Explore a completed Chalk lesson, its accessibility features, and the system behind it. This page clearly presents a saved result rather than a new live render.

Built for learning

Explore a complete lesson Chalk already generated.

General-purpose AI video optimizes for spectacle. Educational video must preserve sequence, legibility, relationships, and timing.

Chalk turns one question into a narrated, hand-drawn lesson. Agents make the teaching decisions; a deterministic compiler enforces the visual rules.

Explore Chalk in three minutes

No account required

Technical explanation

How Chalk works

Creative where teaching needs judgment. Deterministic where clarity needs rules.

  1. Plan01

    Teaching agents make the lesson

    Gemini agents research, outline, edit the narration, establish a visual bible, and direct every scene from one whole-video blueprint.

    Research + outline + parallel scene direction

  2. Compile02

    Typed rules protect clarity

    Bounded scene JSON is sanitized and Zod-validated. TypeScript owns measured text, layout, collision repair, connectors, and speech-led timing.

    Validated scene graph + deterministic geometry

  3. Deliver03

    One render, progressive and downloadable

    SVG frames are deduplicated, rasterized with Sharp, and encoded by FFmpeg into HLS, MP4, captions, and a readable transcript while Convex coordinates queue state and media.

    HLS + MP4 + WebVTT + transcript

The core engineering boundary

A model can propose. Only validated data can render. This keeps the lesson flexible without giving an LLM control over final geometry, timing, or encoding.

Inside the build

What Chalk includes

Each item below describes implemented behavior or an artifact available in the repository. Provider-backed generation is the only feature currently paused.

Problem & solution

Designed for learning, not spectacle

Educational video must preserve sequence, legibility, relationships, and timing. Chalk pairs agentic teaching decisions with deterministic visual rules.

Innovation

Creative planning, deterministic presentation

A model can propose a lesson and its visual intent, but only sanitized, validated scene data can reach the renderer.

Technical achievement

A complete media pipeline

Parallel scene design, semantic icon retrieval, per-beat TTS timing, typed scene graphs, SVG/Sharp rendering, FFmpeg, HLS, and MP4 work as one system.

Architecture & scale

Long-running jobs are isolated from the web tier

Next.js serves the UI, Convex coordinates queue state and media, and outbound leased workers reject stale callbacks. Worker replica scaling is manual today.

UX & accessibility

The explanation stays usable

Responsive layouts, keyboard focus, captions, transcripts, live status, MP4 download, and accessible owner actions cover the journey around the video too.

Completion & confidence

The core workflow is implemented and tested

Deterministic tests, Convex integration tests, desktop/mobile Playwright checks, axe accessibility checks, CI, documentation, and a deployment runbook support the build.

Continue exploring

Explore a real result here, then inspect the implementation in the repository.

This showcase stays available without an active backend or cloud worker, and the saved lesson is always identified as an existing result.