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 · ASPLOS 2027
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.
Co-first author · ATC 2026
Safely extends the eBPF compilation pipeline with native operations while retaining verification of the corresponding eBPF instruction sequences.
Up to 24% faster eBPF microbenchmarks and 12% faster production applications.
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.
A production investigation of DeepSeek V4.1 disk-backed Engram on four DGX Sparks: how syscall, page-cache, and component profiling found a Python scheduling bottleneck, how RWF_NOWAIT fixed it, and what the production evidence does and does not prove.
A complete experiment report on running Qwen 27B with TP2 and native speculative decoding across two RTX 5090 hosts, finding a Thunderbolt transport bottleneck, and validating an opt-in Linux TX-E2E fix.
Community
Talk slides and posters on eBPF, AI agents, runtimes, and systems infrastructure.
ACM SIGOPS ATC 2026, Hong Kong. Paper: github.com/yunwei37/atc26-kinsn-camera-ready; artifact: github.com/eunomia-bpf/bpf-benchmark.
与智能体共同演化的可扩展操作系统 · 第二届 IPADS 青年学术论坛：Agent Harness 专题，2026-10-16
Linux Plumbers Conference 2026, eBPF Track: Mon 5 Oct 2026, 17:00, South Hall 1 A (30 min including discussion).
IntentCap — AgenticOS 2026 lightning talk (5 min)
Attributing agent cost to the work it was doing — AgenticOS 2026 talk (8 min + Q&A)
Verified, Composable Policy Execution for GPU Collective Communication — eBPF '26 talk (17 min + 5 min discussion)
AI × eBPF：面向 AI、由 AI 改进的可编程运行时与操作系统
Extending eBPF Beyond Kernel Extensions Verified Interfaces for Runtime System Extensibility
Fork, Explore, Commit: OS Primitives for Agentic Exploration Agentic OS Workshop, ASPLOS 2026
ACRFence: Preventing Semantic Rollback Attacks in Agent Checkpoint-Restore CoDAIM 2026 Workshop
AgentCgroup AI Agent Resource Characterization and OS-Level Control
OpenClaw Open-Source AI Agents: Personal Assistants, Security, and Open Research Questions
OpenClaw & AgentCgroup Open-Source AI Agents: Personal Assistants, Resource Characterization, and OS-Level Control
Extending eBPF to GPU Device and Driver Contexts Linux Plumbers Conference 2025 - eBPF Track
Towards Agentic OS: An LLM Agent Framework for Linux Schedulers MLforSystem 2025 Workshop @ NeurIPS
SchedCP: Autonomous OS Optimization with LLM Agents A framework for safe, efficient, and autonomous performance tuning.
GPT-OSS (20B, 120B) & the "Safety Tax" Open-Source Reasoning Models and the Cost of Safety