Working on systems where algorithms, compilers, and hardware are designed as one.

I'm broadly interested in deep learning systems, compilers, and hardware-aware programming abstractions. The two questions I've been thinking about recently are:

• What happens when the algorithm, software, and hardware layers are designed together, rather than stacked on top of each other?
• What does it take for large-scale human–LLM collaboration — across people and across agents — to sustain delivery inside a real software stack over the long run?

Outside of research, I enjoy writing programs and building software systems — I just like making things.

Both projects are joint work with friends at @tile-ai.

An ongoing effort that grew out of our earlier research — exploring what the layers between algorithms and hardware should look like when they are designed together, rather than stacked on top of each other.

An exploration of how far LLM agents can go in autonomously developing an operator library — from writing kernels to testing and iterating on them — with quality good enough to actually ship.

• 🚀 TileFusion — an experimental C++ macro kernel template library that raises the abstraction level of CUDA C for tile processing, so algorithm developers can innovate on hardware-aware LLM kernels without drowning in low-level details.
• 🧩 FractalTensor — a programming framework built around FractalTensor: nested, statically-shaped tensor lists with functional array operators (map / reduce / scan). DSL + IR work inspired by polyhedral loop analysis. [paper]
• 🔍 VPTQ — an extreme low-bit quantization algorithm and inference library for LLMs, led by my friend @YangWang92; I contribute on the systems side.

I keep a blog where I jot down ideas that catch my attention in daily work — updates are infrequent but unhurried. @haruhi55 is also me in disguise. 🐵✨

lcy.seso@gmail.com · caoyingseso@126.com

Feel free to reach out — happy to talk about deep learning systems, compilers, hardware co-design, or LLM-driven engineering.