Hi, reader! I am a fifth-year CS PhD student at Stanford. My advisor is Alex Aiken. I have done two (extended) internships with NVIDIA's Programming Systems and Applications Research Group, managed by Michael Garland.
Want to talk about your research? I am organizing the Stanford Software Research Lunch this year. Please email me with a talk proposal and I will get back to you.
Optimizing the inference stack
My main research is on compilation and formal methods for the inference stack.
Currently, I am working on optimal routing algorithms for distributed inference systems like Dynamo. More details soon.
Twill is the first optimal approach to software pipelining and warp specialization for GPU kernels. Twill has been reimplemented at several places including Meta and CMU.
Weft++ (working title) is a sound and complete deadlock- and race-checker for GPU kernels. A preliminary manuscript is available.
Equality saturation
A separate line of work focuses on equality saturation (EqSat), a new technique for building program optimizers.
My collaborators and I performed the first extensive evaluation of EqSat against another program optimization technique. We are continuing this effort to produce a standard benchmark suite and competition format for equational program optimization.
I have also worked on applying EqSat to instruction scheduling and developed an incremental version of EqSat.
High-performance computing
Early in my PhD and during my undergrad, I worked on the Legion parallel programming system and its domain-specific language Regent. This work resulted in two papers, my undergraduate thesis, and several PRs to Regent.
Other Writing
- How to See Art new Jul 2026
- The Next 700,000 Programming Systems May 2026
- VLIWs, and How to Schedule Them Mar 2026
- Compilers Should Have a Superoptimize Flag Dec 2025
- The Unspoken Expectation Dec 2024