Founding Engineer, Outpaint (YC W21)

September 2024 — Present New York, NY

Outpaint (formerly True3D) is a research lab building immersive media: capture, compression and playback of 3D scenes that you can move around inside. I was the first engineer hire and own the rendering and media processing end to end: the codecs that get 3D data onto the GPU, the renderers that draw them, and the training code that produces them in the first place. Everything has to run in a browser tab on a laptop, so I was deep in both CPU and GPU optimization.

Highlights

  • Built a voxel-video decoder in WebGPU that is ~3× faster than the WebAssembly decoder it replaced. Implented a block-parallel entropy decoder and a work-efficient prefix sum, hitting 10B voxels/s in-browser on consumer laptops.
  • Shipped two WebGPU voxel renderers: raycasting for opaque scenes, splatting for transparent ones. They render up to 1B voxels at 60fps (1080p, M4 MacBook). These powered the interactive investor demos for our seed raise.
  • Wrote a CUDA differentiable renderer for training 4D Gaussian splats with custom attributes (polynomial motion, time-constraint), producing models 5× smaller than per-frame 3DGS.
  • Developed a novel UX paradigm for viewing 3D media on a flat screen using iris tracking via MediaPipe. Open-sourced the implementation for 350 GitHub stars.
  • Designed a spatial-photo format that re-represents splat output as mesh + atlas geometry: >99% smaller, natively rasterizable, no per-frame splat sorting.
  • Built a WebGL depth-video renderer and WebXR viewer, using WebCodecs for frame-accurate depth/color sync.

Projects here

  • SPLV (Voxel Video Codec)

    2025

    A first-of-its-kind 3D video codec based on voxels. Both WASM and WebGPU-accelerated decoders on the web.

    • WebGPU
    • WGSL
    • C
    • WebAssembly
    • Compression
    • GPU parallelism
  • 4D Gaussian Splats

    2025

    A differentiable CUDA renderer for training 4D Gaussian splats, and a WebGPU renderer for in-browser viewing.

    • CUDA
    • PyTorch
    • C++
    • WebGPU
    • Differentiable rendering
    • Gaussian splatting

Media

A "Spatial Photo" - generated from a single regular photo.
A demo of the iris-tracked 3D video viewer.
A 4D Gaussian splat, trained using my differentiable CUDA renderer. Rendered using my WebGPU renderer.
1B voxels decoding and rendering at 30fps in Chrome, 1080p on an M4 MacBook.
Photo of the Outpaint team
The team. Photo taken after our first ever sale.