Research

My work sits at the boundary between compilers and hardware. I build systems that reason formally about runtime behavior at compile time: a profiler that measures how long data actually lives inside an accelerator, and a compiler framework that turns those measurements into a provably optimal placement schedule across heterogeneous memory. The systems themselves are described on the Projects page; this page covers the research context and the publication record.

Memory Systems for AI Accelerators: LtRAM and StRAM

September 2024 – March 2026

For the first part of my Ph.D. my primary research direction was memory systems for domain-specific hardware accelerators, in collaboration with researchers across the Computer Science and Electrical Engineering departments at Stanford. This work was carried out within the Differentiated Access Memories (DAM) project, a multi-year Stanford effort to define how future computers will manage memory as a heterogeneous resource with different tradeoffs. My involvement in that project ended in March 2026.

I was interested in the “memory wall” problem in the AI and ML fields, where the surging size of AI models demands more on-chip memory to store intermediate results. Even as logic continues to scale down with more advanced process nodes, the density of SRAM has not kept pace, creating the risk that memory comes to dominate the area and power budgets of future AI accelerators. That points to renewed interest in alternative on-chip memory devices offering higher density and better fJ/bit efficiency.

We argued that memory will evolve from a uniform address space into a heterogeneous collection of devices optimized for different uses. The questions this direction sought to answer: what new memory technologies will we need, how do we compose them in systems, and what caching mechanisms do they need? How can applications and algorithms guide the composition of heterogeneous memories? How will software present and manage them?

My contributions to this direction, chronologically:

• GainSight: given 2T/3T gain cells as denser SRAM alternatives with limited retention times, I built the first comprehensive, open-source framework that aligns dynamic workload lifetime profiles with memory device characteristics to generate optimal on-chip memory compositions, making the design space of heterogeneous on-chip hierarchies something you can measure rather than guess at.
• Position papers on the Future of Memory (BigMem '25) and Towards Memory Specialization (DIMES '25): two co-authored papers arguing for a shift away from uniform memory hierarchies toward specialized architectures that exploit application-specific access patterns and data lifetime characteristics, proposing long-term RAM (LtRAM) and short-term RAM (StRAM) as new memory classes with explicit OS support.
• OpenGCRAM: contributing to an open-source gain cell memory compiler for generating GCRAM designs with customizable retention times for AI workloads.
• Remi: a compiler analysis framework that jointly optimized scheduling and retention-aware memory planning for AI accelerators with short-term RAM scratchpads, minimizing refresh overhead by exploiting per-row retention variation. The proposal was a Qualcomm Innovation Fellowship 2026 finalist.

Current Direction

Fall 2026

Through the Fall 2026 academic quarter I am rotating with Prof. Keith Winstein, reading into deterministic and content-addressed computation, distributed dataflow and task-graph execution, and the formal semantics of program execution. I am interested in how deterministic dependencies, memoization, and precise evaluation rules might support operator fusion, symbolic execution, and termination prediction at the compiler level.

Undergraduate-level Research

Most of my research experience prior to Stanford was at the University of Michigan, through the Civil and Environmental Engineering department and the University of Michigan Transportation Research Institute (UMTRI).

Road-side based cybersecurity in connected and automated vehicle systems, January 2022 – April 2024

This project encompassed my research work done under the guidance of Prof. Neda Masoud at the University of Michigan Center for Connected and Automated Transportation. The overarching objective of the project was “to develop a holistic framework that integrates physics-based data-driven modeling and dynamic decision making under uncertainty and partial information to improve cybersecurity in connected and automated vehicles (CAV).”

UMTRI Bioscience Group, February 2022 – August 2022

I worked as a research assistant at the University of Michigan Transportation Research Institute's Biosciences Group over the Winter, Spring, and Summer 2022 terms. The group “conducts research on the biomechanics of motor vehicle occupants as it relates to occupant injuries, crash protection, and occupant accommodation.”

Publications, Preprints and Conference Presentations

You can also find my articles on my Google Scholar profile.