- 1M+Lines of code
- 500+Renders with early partners
- 8KMaximum output resolution
Videomath started in 2019 as a personal Python library. I had just started college, was watching a lot of 3Blue1Brown, and wanted to make my own explainer videos for YouTube. Over several rewrites of the core engine it grew into a full procedural video system.
How it works
Rather than generating pixels with a diffusion model or throwaway web code, Videomath builds video from computer graphics primitives: shaders, grid and camera systems, and scientific tooling for 2D and 3D charts. It also includes chemistry and electronics (SPICE) simulation engines, geographic datasets and world maps, Blender integration and headless web rendering.
Because the engine was designed for people first, its code is organised and expressive. That turns out to be ideal for AI agents: a short prompt becomes a small amount of code, and the engine handles research, art direction, narration, score, pacing and an internal critic that reviews renders before they ship.
Where it is now
Videomath runs in invite-only early access with partners, who have produced more than 500 renders so far. Output ranges from seconds to hours, in 16:9, 1:1 or 9:16, up to 8K.