MS CS @ USC (May 2027). I work on LLM inference and AI infrastructure: serving systems, KV-cache movement, and the performance work that makes them fast. Before grad school, 7 years of software engineering at Samsung and Continental.
Open source contributions: vLLM
- #57648 (merged). Removed the
ray[default]dependency from data-parallel placement-group setup, which broke elastic expert-parallel scale-up on Ray clusters without the dashboard. - #54483 (merged). Coalesced NIXL host-buffer KV-cache copies across cache groups. Per-request copy latency 4.95ms to 1.27ms, device copy ops 96 to 16 at 6 KV-cache groups.
Recently
- Amazon, SDE intern (2026). Productionized prefill-decode disaggregation for LLM serving, and root-caused per-request KV-transfer variance down to the RDMA post stage.
- Second author on a paper evaluating how well LLMs reproduce community reactions to news, under review at ICLR 2027.
Projects: EchoSlice, LLM-backed TED clip recommendation and spaced repetition.