ML Systems · GPU Runtimes · Programmable Systems
Ph.D. Candidate in Computer Science and Engineering at UC Santa Cruz
Building programmable and observable runtimes for AI infrastructure.
My research spans GPU runtime and policy mechanisms, distributed-training performance diagnosis, Linux extensibility, and system support for autonomous agents. I am expected to graduate in 2027 and am a Research Scientist Intern (AI Infrastructure) at ByteDance in Summer 2026.
Expected Ph.D. 2027 • Santa Cruz · San Francisco Bay Area
Research
Systems are the primary unit of my work: each combines a concrete mechanism, an implementation, and evaluation or downstream use.
Co-first author · research design & prototype
Continuous cross-layer diagnosis for production AI training, combining CPU stack profiling, GPU kernel tracing, and NCCL instrumentation with less than 0.4% overhead.
Validated by Alibaba collaborators across 80K+ GPUs; 94 confirmed production issues diagnosed.
Creator / lead maintainer · OSDI 2025
A userspace eBPF runtime and verified application-extension framework combining verification, isolation, and dynamic instrumentation without application recompilation.
1.5K+ GitHub stars, active external users, and downstream integrations in the eBPF ecosystem.
First author · GPU runtime research
A programmable policy runtime spanning the Linux GPU driver and device, with verified hooks and device-side execution for scheduling, memory, and observability policies.
Up to 4.8× throughput improvement and 2× lower tail latency on evaluated AI workloads.
Creator / lead maintainer · agent runtime observability
System-level boundary tracing that correlates LLM and agent activity with process, file, network, and OS effects without requiring application SDK instrumentation.
Less than 3% average overhead; integrated into Alibaba Cloud Linux 4 Agentic Edition (ANOLISA).
Writing
Research notes, systems experiments, and occasional writing outside computer systems.
Community
Technical talks and community presentations on eBPF, runtimes, and systems infrastructure.
ACRFence: Preventing Semantic Rollback Attacks in Agent Checkpoint-Restore CoDAIM 2026 Workshop
View slides →
Extending eBPF Beyond Kernel Extensions Verified Interfaces for Runtime System Extensibility
View slides →
AgentCgroup AI Agent Resource Characterization and OS-Level Control
View slides →