Computer vision that respects your privacy

Why build this?

Browsers have quietly become a capable ML runtime. WebAssembly, WebGL compute, WebGPU, SIMD and worker threading mean a laptop running Chrome can do what needed a dedicated GPU five years ago. Benzo Vision is an opinionated tour of that stack: a real product, not a demo, that stays 100% on-device.

Tech stack

• MediaPipe Tasks Vision — object detection, pose, hands, face
• WebAssembly SIMD — CPU fallback for when no GPU is available
• WebGL 2 — GPU delegate for inference kernels
• IndexedDB — model blob cache + user preferences
• Canvas 2D — overlay annotation renderer
• MediaRecorder — annotated video export
• Vanilla ES modules — zero build step, zero dependencies

Engineering principles

• No network after first load. Every model is cached, so the app works offline after the initial visit.
• Never block the UI thread. Inference happens at a capped rate; the compositor drives the render loop.
• Prefer the smallest usable model. Lite variants ship by default; heavier models are opt-in.
• Compose, don't couple. Modules talk over a shared event bus — the engine doesn't know about the UI.

Roadmap

Credits

Built with MediaPipe by Google. Model licenses are Apache-2.0. Everything else is MIT.