Youngwook Paul Kwon

YW Paul Kwon

young at berkeley.edu

About

Building something [un]conventional.

My past work spans both foundation model infrastructure and applied perception systems: at SambaNova, I built infrastructure to enable the latest LLMs on specialized AI hardware, and at Phantom AI, I developed 2D and 3D object detection systems for autonomous driving.

I graduated with my MS and PhD from UC Berkeley and a dual degree from Seoul National University.

Experience

  • Member of Technical Staff, Unconventional AI [04/26 - present]

  • Senior Principal Engineer, Sambanova Systems [05/24 - 04/26]

  • Staff Deep Learning Engineer, Phantom AI [05/17 - 05/24]

  • Research Intern, Lawrence Livermore National Lab [12 - 16 summers]

  • Research Assistant/Teaching Assistant, UC Berkeley [08/11 - 05/17]

Publication

full / selected * equal contribution

Patents

Projects (curiosity, side, fun)

Continuous Thought Machine Lab
Continuous Thought Machine Lab
Experimental research on dynamics-based neural computation.
Built training + evaluation pipelines and ran comparative experiments vs standard architectures.
Siamese Network Implementation using Tensorflow on MNIST
Siamese Network Implementation using Tensorflow on MNIST
Simple Siamese network with 2D embedding visualization.
Created clean educational implementation (250+ stars).
yank-path.nvim
yank-path.nvim
Minimal Neovim plugin for copying file path variants.
Built lightweight Lua plugin with zero dependencies.
Document N spliter
Document N spliter
Deterministic document segmentation via atomic parsing.
Designed and implemented structured parsing + partitioning system.
Delaunay Triangulation Implementation
Delaunay Triangulation Implementation
2D Delaunay triangulation from scratch.
Implemented incremental Delaunay with BBCK star graph and randomized point location; visualized in OpenGL/Python.
video
im2hyperdrawing
im2hyperdrawing
Computational photography for time-lapse drawing synthesis.
Built image-to-drawing pipeline using Poisson blending, mean-shift segmentation, and Canny edges; renders as a time-lapse drawing video.
twix