|
Research Interests
I work at the boundary between intelligent software and the systems that make it practical, with particular emphasis on:
- Systems for AI: efficient, reliable runtimes for agentic and generative workloads
- Heterogeneous Edge AI: correctness-aware placement across CPUs, iGPUs, and NPUs
- Compiler and Runtime Optimization: reducing compilation, memory, and state-migration costs
- Program Analysis and Systems Security: turning software fixes into reusable bug detectors
|
|
Research Projects
My current work spans edge AI systems and automated systems-security analysis. Representative projects are
highlighted; see also my
Google Scholar profile.
|
|
JIT Specialization for Static-Shape NPUs
Vikranth Srivatsa, Xuming Huang, *Prof.
Yiying Zhang
WukLab @ UC San Diego — 2026 · *advising faculty
tech blog /
full experiments /
campaign data /
compiler patch /
PR branch /
OpenVINO NPU Compiler
A correctness-first runtime that grows an LLM through 1K→2K→4K→8K NPU specializations while the CPU
compiles exactly one step ahead. On NUC16, five counterbalanced pairs cut foreground compile wall time by
22.49% and the measured protocol critical path by 8.37%, with 15/15 successor deadlines met and 20/20
boundary outputs matching the eager control. Separately, compiler tracing and memoization reduced full
compile time by 3.34–3.68% in 12/12 A/B trials.
|
|