Equality saturation at scale
Foresight makes e-graphs parallel and programmable; Latent Idiom Recognition uses equality saturation to recover hidden high-performance library calls.
AI Accelerators · Compiler IRs · Equality Saturation
I build compiler systems for accelerators, FPGAs, managed runtimes, and MLIR tooling. My work focuses on IR design, hardware-aware compilation, and optimization techniques including equality saturation.
Now
I recently joined Cerebras as a compiler engineer, where I work on compiler infrastructure for AI accelerator systems. Alongside that, my PhD research at McGill University and Mila, with Christophe Dubach, turns equality saturation and compact functional IRs into tools for parallel rewriting, FPGA design search, and library idiom recognition.
Guide
Foresight makes e-graphs parallel and programmable; Latent Idiom Recognition uses equality saturation to recover hidden high-performance library calls.
SkeleShare searches hardware resource-sharing designs with equality saturation, solver-based extraction, and hardware-conscious program structure for FPGAs.
MLIR.NET, Flame, and the Julia GPU garbage collector cover typed IR APIs, SSA optimization, LLVM lowering, managed runtimes, and GPU runtime support.
Featured Systems
CC 2026
A parallel, extensible equality saturation library with programmable saturation strategies, generalized e-graph metadata, and deferred parallel rewriting.
CGO 2026
Automates hardware resource allocation and sharing by combining equality saturation, algorithmic skeletons, and solver-based extraction.