Research
My work is centered on AI systems and computer architecture. I am interested in how compute, memory, and software interact during language-model inference — and in the design choices that make serving large models faster and more predictable.
Three directions connect my work: accelerator and memory architecture, inference-system design, and the algorithms that shape execution.
Research Interests
- AI accelerator architecture — memory-centric accelerator design for large language models, including 3D DRAM-based architectures.
- LLM inference & serving systems — speculative decoding, draft-model architectures, and serving efficiency.
- GPU systems & kernel execution — persistent kernels and system-level execution for GPU workloads.
- Agent systems & empirical LLM theory — tool-using agent workflows and measurement-driven studies of model behavior.
Current Projects
Exploring memory-centric hardware for language model acceleration.
Architectures aimed at improving speculative token acceptance length.
Persistent GPU kernel research at CMU Catalyst Lab.
Descriptions above reflect ongoing project topics only. Unpublished measurements, confidential details, and code from industry collaborations are intentionally omitted.