yl@lab ~ % python sample.py --prior gaussian
x0 ~ N(0, I)
integrating dx/dt = v(x, t)

YL-26

林煜轩 Yuxuan Lin

Computer Engineering @ Columbia

I care about bringing AI models into real systems that are usable, understandable and scalable. My engineering work spans multi-agent applications, distributed training, CUDA optimisation and embedded robotics.

The name on the screen is written by particles: they start as Gaussian noise and follow a velocity field to the glyphs, the way Flow Matching turns noise into data. Push the t slider on the machine to see every step in between.

Now: Columbia UniversityM.S. in Computer Engineering
Undergraduate: Zhejiang University × UIUCZJU–UIUC dual-degree program
Paper: IEEE ISBI 2026Second author

The same kind of growth

lightning, rivers, blood vessels, leaf veins, snowflakes, binary trees, galaxies, the DNA double helix and point-cloud terrain: one family of fractal structures recurring at every scale. Each is drawn by a few lines of maths.

Work

All 6 projects

Writing

All 42 pieces (Chinese)

Start from a question

My notes and articles are written in Chinese.

Why are GPUs fast?CUDA
How is a model trained across GPUs?Distributed Training
How do generative models transform distributions?Flow-based Model Road to Flow Matching
How do you build an AI application?RAG LLM Evaluation