Designing the machines that think.
I work on machine learning systems end to end: models and inference at scale, the compilers and kernels beneath them, and the accelerators they run on. Most of all I care about designing those layers together.
Notes & derivations
Codesigning an Accelerator from Agent Traces
From a profiled Claude agent session to a custom accelerator: hardware models, an area-allocation proof, the Amdahl wall, and a 4× projection.
Understanding the SemiAnalysis Inference Curves
Rooflines, the KV-cache tax, TTFT and inter-token latency: a model for reading inference throughput frontiers.
Lectures & tutorials
- High Performance Matrix Multiplication on RDNA3 GPUs Sep 2023 Slides Video
- High-dimensional layouts for vector distribution May 2023 Slides Video
- Decomposable Operators in IREEC4ML Feb 2023 Slides
- Optimizing Winograd Convolutions in MLIRMLIR Summit Nov 2022 Slides
- MLIR Code Generation TutorialHot Chips 34 Aug 2022 Slides Video
Published work
- Wave: A Symbolic Python DSL and Compiler for High-Performance Machine LearningMenon, Zinenko, Verma, et al. — MLSys 2026PDF
- Weighted Basis Images: Learning Alternative Data Representations 2021PDF
- Deep-Learning based Approaches to Visual-Inertial Odometry for Autonomous Tracking ApplicationsMenon, Ramachandrappa, Kesinger — SID Symposium Digest 2018DOI
- Adjoint-based Aerodynamic Shape Optimization Using a Spline-Based Parametric Geometry Framework 2012PDF
- Feasibility Study of a Unified CFD-CSD Computational FormulationMussfar, Sankaran, Sitaraman, Menon, Farhat 2010PDF
Where I’ve worked
-
Machine learning systems across the whole stack: how models and inference map through compilers and kernels onto AMD hardware, and how those layers should be designed together.
-
Built the company’s technology end to end: the infrastructure that took machine learning models from frameworks to fast execution on a wide range of hardware. Acquired by AMD.
-
Autonomous systems for electric vertical takeoff and landing aircraft.