I am a researcher and engineer working across LLMs, multimodal AI, and ML systems. I recently graduated from Georgia Tech with dual M.S. degrees in Computer Science (Machine Learning) and Computational Science & Engineering (Applied Mathematics), and I am currently a Research Intern at Stanford's CogT Lab, where I work on efficient adaptation and knowledge transfer for foundation models.
My research focuses on foundation model reliability and efficient AI systems. I study how large language models and vision-language models use evidence, revise decisions, represent uncertainty, and behave as inference-time compute scales. Recent work spans multimodal RAG, self-correction, RLVR, uncertainty estimation, efficient VLM inference, and visuomotor learning. I am particularly interested in settings where stronger models, additional context, or more compute do not automatically translate into more reliable behavior.
On the systems side, I build GPU-optimized inference pipelines, large-scale evaluation infrastructure, and production serving systems for generative models. Previously at GMI Cloud, I optimized Flux-Schnell (12B DiT) inference on H100 GPUs to ~30 images/min at 1–2s latency and worked on multi-GPU inference and serving with TensorRT and NCCL.
I also contribute to open-source ML infrastructure, with bug fixes, tests, and systems improvements merged upstream into Hugging Face Transformers, PEFT, timm, Datasets, and Accelerate, as well as PyTorch TorchRL and TensorDict, Google ADK and gemma.cpp, LinkedIn Liger-Kernel, NumPyro, and Axolotl.
I grew up in China and did my undergraduate degree in Artificial Intelligence at Shandong University. Those four years gave me a strong grounding in mathematics and control theory, and more importantly, taught me how to think across disciplinary boundaries. I have carried that habit with me ever since.