Xiakai Pan (潘夏凯)
AI Systems Research · Computer Architecture · Compiler/Runtime · Agent Systems
Research Statement · Research · Publications · GitHub
Contact: xiakpan@gmail.com
Research Interests
Structural modeling, representation and transformation for heterogeneous AI computing; workload/system scheduling, compilers and performance modeling; long-horizon Agent execution; formal task evolution.
Selected Research
SSG — System Substrate Graph
Workload–System–Schedule representation and Build–Traverse–Translate mechanism for heterogeneous AI execution, including operators, training/inference/RL, and Agent/tool/model workloads. The former Agent-SSG work is integrated into SSG; it is not a separate current paper. A 2026 SSG Technical Report — author edition (not peer-reviewed) reports bounded H20 and 8×A800 results and negative correctness/fallback evidence. Later generalizations require separate validation.
TSG — Task Substrate Graph
Formal study of persistent, evolving human–AI tasks using Actor × Spec × Impl, semantic context, activity-relative Component/Model/Artifact roles and intra-/inter-task evolution. 2026 self-hosted preprint; evaluation and semantic questions remain open.
SSG-DT — Predictive Execution Modeling
Research direction on structure-preserving cost/simulation models over selected SSG execution plans. Design and evaluation protocol available; no validated twin results claimed. My Three Models note discusses formal, simulation and statistical modeling roles.
Peer-reviewed Publication
Chen Li, Xiakai Pan, et al. Style Adaptation module: Enhancing detector robustness to inter-manufacturer variability in surface defect detection. Computers in Industry (2024). DOI.
Research and Professional Experience
Baidu — Heterogeneous Computing Engineer
September 2025 – Present
- GPU/NPU inference execution infrastructure, operator implementation and runtime adaptation.
- Accelerator operator integration with vLLM, including AWQ-related work.
- Research prototypes for cross-scale execution representation and scheduling.
Li Auto — NPU Architecture and Modeling Engineer
July 2024 – September 2025
- Accelerator functional/performance modeling; on-chip interconnect, vector computing and memory hierarchy studies.
- gem5-based exploration, workload-aware hardware/software co-design, and compiler/runtime integration.
Borui Jingxin — NPU Architecture Modeling Intern
November 2023 – January 2024
NVDLA-oriented simulation infrastructure and model integration.
PLCT Lab — RISC-V Virtualization Research Intern
July 2022 – January 2023
RISC-V hypervisor extensions, KVM/QEMU/Spike, and a ten-article technical series on RISC-V virtualization published on TinyLab in 2023.
Education
Tsinghua University — M.S., Data Science and Information Technology, 2021–2024; computer architecture, CPU simulation and industrial vision.
Nankai University — B.S., Chemistry, 2017–2021; additional study in Chinese language and literature and experience in machine learning.
Technical Writing
RISC-V Virtualization: 10 technical articles, TinyLab / 泰晓科技 (2023). Covers privilege transitions, Host/Guest boot, memory virtualization, KVM userspace, exception/interrupt handling, and Spike/QEMU/kvmtool implementations. Complete series and original links.
Technical Skills
C/C++ · Python · CUDA/Triton · GPU/NPU operators · TVM/MLIR/XLA · vLLM · gem5/SST/QEMU/Spike · functional/performance modeling · AI inference and distributed scheduling.
Updated November 2026. Statements of research scope do not imply completed implementation. Experimental figures are attributed to the linked manuscript and have not been independently reproduced for this website.
Typeset copy
The typeset PDF is not published here. If you need it for a specific purpose, write to xiakpan@gmail.com and say what it is for; I will send the current version directly.