My research spans the inference and training stack: deciding when a small local model's answer is good enough before paying for a frontier one, packing training workloads onto partitioned GPUs, and simulating wafer-scale interconnects so runtime changes can be validated in minutes instead of days.
I studied Computer Engineering at Columbia University (B.S. 2026), where I published first-author work on GPU partitioning and fine-tuning efficiency with Prof. Martha Kim and IBM Research. At Cerebras Systems, I designed and built a simulator of the wafer-scale engine's interconnect from scratch — now part of how runtime engineers there diagnose throughput stalls on large models.
Preprint, 2026
Generate locally first, then decide. A lightweight probe over the local model's hidden activations scores each response (ROC-AUC 0.861 at detecting correctness) and escalates to a remote model only when the local answer looks insufficient, matching always-remote accuracy while escalating just 54% of queries.
Draft PDF Available on Request
IEEE International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2025
Treats a single MIG-partitioned GPU as a multi-objective scheduling problem under real diurnal datacenter load, with reinforcement-learning-driven repartitioning — 68% better on a joint energy–tardiness objective than an unpartitioned GPU.
Machine Learning for Computer Architecture and Systems (MLArchSys, at ISCA), 2025
A profiling study of LoRA rank on Llama 3.1 8B across throughput, energy, and memory, showing up to 2.7× efficiency gains with negligible impact on perplexity.
Workshop on Machine Learning and Systems (EuroMLSys, at EuroSys), 2024
Measured six workloads, from 14M-parameter classifiers to 1.5B-parameter language models, across MIG partition mixes: heterogeneous partitions deliver up to 33% lower energy and 9% higher training throughput from a single GPU, and fine-tuning runs 55% faster at 42% less energy.
Earlier work, before ML systems: regenerative medicine at Penn's Singh Center for Nanotechnology (2021), and radio astronomy — modeling Jupiter's synchrotron radiation with NASA JPL collaborators (ATOM, 2020) plus two star-cluster studies (SAS, 2020).
Selected engineering work, from silicon to kernel to runtime.
Before this existed, validating a runtime change meant a 24-hour run on wafer hardware. I designed and built a C++ simulator of the wafer's inter-die fabric from a blank file: every region, core, and wire, with backpressure from link capacities and delays, so simulated throughput and stalls track the real machine. Iteration dropped to 5 minutes (288x), and the simulator surfaced KV-cache token drops, batching inefficiencies, and race conditions on production-scale models. Hardware, systems, and ML teams adopted it, and engineers I onboarded kept extending it after my internship ended.
Extended the HLS4ML compiler with custom C++ translators for matmul and transpose, implementing the hardware primitives transformer attention needs on FPGAs within the open-source ESP SoC platform. Report
Hardware/software co-design in SystemVerilog and C running an int8-quantized TensorFlow CNN (scales and zero-points handled in hardware), reaching a 3.4x convolution speedup at 30% device utilization. Presentation
A custom priority scheduler (+75% throughput on mixed workloads) and a complete read/write filesystem, both implemented in-kernel in C. Notes
Real-time control software coordinating ten lab devices, including manipulators, pumps, and cameras, to 1.0 µm precision for automated patch-clamp electrophysiology. GitHub
Where this started.
One of 350 students worldwide selected for a physics-based game simulating collisions and gravity, written in Swift. Apple flew me to San José as a high-school freshman. Code
An iOS ordering app for my high school's coffee bar — Stripe, Apple Pay, CloudKit, and AWS Lambda — that ran in production on the App Store through two years of daily lunch rushes.
Real-time object tracking running on an iPhone mounted to a competition robot, streaming target poses over TCP to the drive controller. GitHub
Also from those years: augmented-reality turn-by-turn navigation prototypes, and websites built and delivered for local clients.