I am Kan Zhu, a fourth year PhD student at University of Washington’s Paul G. Allen School of Computer Science and Engineering, co-advised by Baris Kasikci and Arvind Krishnamurthy.

I build systems that make large language model (LLM) inference faster and cheaper. My work spans the inference stack: at the algorithm level, I reduce computation and memory traffic with sparse attention and low-bit quantization; at the engine level, I design serving systems and schedulers that make full use of GPU and CPU resources. I am now extending this work to agentic AI systems, whose multi-step, tool-using workloads change how inference engines should schedule requests and reuse computation.

Download my CV