CW

Training and inference systems for large-scale AI.

I am an MTS on Training at Fireworks AI. Previously, I did my Ph.D. work in Computer Sciences at the University of Wisconsin-Madison, advised by Prof. Suman Banerjee. My work focuses on scalable training infrastructure, RL-based post-training, high-throughput long-context inference, and systems-agent co-design.

Training Systems RL Post-training Long-context Inference Cache-aware Agents

2026/09 - Present

MTS, Training · Fireworks AI

Working on training systems for large-scale AI workloads.

2026/05 - 2026/09

SWE PhD Intern · Google

Engineered straggler injection, profiling, and diagnosis workflows for NCCL FasTrak across A3 Ultra and A4 GPU clusters.

2025/05 - 2025/08

SDE Intern · Meta

Developed production ML pipelines for ads retrieval and built LLM evaluation workflows for retrieval-quality regression tracking.

2024/05 - 2024/08

Research Intern · ByteDance

Optimized PCIe GPU collective communication and studied RDMA scalability bottlenecks for expert-parallel LLM inference.

2023/05 - 2023/08

Research Intern · MSRA

Improved volumetric-video inference with LUT-based acceleration and drafted work on efficient 3D point-cloud enhancement.

Selected research output

View all
ViTL paper thumbnail

ICML 2026 · Long video QA

Video-in-the-Loop: Span-Grounded Long Video QA with Interleaved Reasoning

Long-context video reasoning with temporal grounding, evidence localization, and interleaved inference.

Zoomer paper thumbnail

TMLR · Inference systems

Zoomer: Enhancing MLLM Performance with Adaptive Image Focus Optimization

Focus-aware image optimization for stronger model quality and efficiency.

SAN architecture paper thumbnail

NSDI 2026 · Storage systems

Switched or Switchless: An Empirical Study of SAN Architecture for Disaggregated Storage

Empirical systems study of SAN architecture choices for disaggregated storage.

VoLUT paper thumbnail

MLSys 2025 · Streaming systems

VoLUT: Efficient Volumetric Streaming Enhanced by LUT-based Super-resolution

Learning-enhanced volumetric streaming system for immersive content delivery.

Notes and blog posts

To be determined

Writing will live here once the blog is ready.

Outside work

I enjoy basketball and traveling. I am also a fan of Shanghai Shenhua.