research

3D vision, GPU systems, and AI agents.

Remote GPU acceleration & cross-platform graphics

HKUST SPARK Lab & WiCi AI · July – September 2026
Undergraduate Visiting Research Student & R&D
Supervisor: Prof. Zili Meng

I worked on keeping data and code local while running demanding GPU workloads remotely over commodity Wi-Fi. This included adapting four 3D Gaussian Splatting frameworks and developing remote graphics support for macOS applications.

  • Remote 3DGS training. Reached 83.28% of native training-loop throughput with gsplat, and 82.27% training-loop / 78.11% end-to-end efficiency with Original 3DGS. The reported comparison achieved a 3.26× speedup over a local RTX 4060 with no quality degradation.
  • Metal-to-Vulkan graphics. Worked on a compatibility layer with remote rendering and video streaming. GravityMark 1.89 at 1080p, one million particles, and Temporal AA averaged 45.8 application FPS, compared with 7.7 FPS on the MacBook Neo’s native Metal path.
  • DaVinci Resolve workloads. Ran selected Metal workloads on a remote RTX 5090 Laptop GPU. Exporting 500 frames of Standard Candle UHD: Rocket Science (66 Blur), at 1080p / 25 fps / AV1, took 6m 08s, compared with 13m 33s natively on a MacBook Pro M5 Max.

research-harness

Independent project, Zhejiang University · April 2026 – present
Supervisor: Prof. Yiyi Liao · Code on GitHub

I am developing a multi-loop agent framework for long-horizon research. An inner loop turns self-review into targeted revisions, while an outer loop uses interface contracts and execution tracebacks to repair the workflow.

The architecture separates task-specific, pluggable skills from reusable orchestration and recovery logic.

Segmentation-aware 3D Gaussian Splatting

Student Research Training Program, Zhejiang University · March 2025 – May 2026
Project Lead · Supervisor: Prof. Jixing Cao

I explored how segmentation can focus reconstruction on the foreground, allocating the Gaussian budget to relevant regions rather than the background.

  • Foreground-focused reconstruction. In the reported RTX 4060 experiment, training time fell from 142 to 61 minutes (56.7%), L1 loss fell by 75.2%, and PSNR improved by 0.57 dB.
  • Sparse-view augmentation. Built a hybrid 3DGS–diffusion workflow to augment 8–12 input images before training, and evaluated reconstruction across scenes with varying structural complexity.