Education
- Major: Physics (Course 8), Artificial Intelligence and Decision Making (Course 6-4)
- GPA: 5.00/5.00 (as of Jun. 2025)
- Selected Coursework: Machine Learning, Deep Learning, Natural Language Processing, Design and Analysis of Algorithms, Computational Architecture, Probability and Statistics, Quantum Physics I, II & III, Physics of Solids, Statistical Mechanics II, Quantum Field Theory I
- GPA: 3.91/4.00
- Selected Coursework: Introduction to Computation, Data Structure and Algorithms, Classical Mechanics, Quantum Mechanics, Thermodynamics and Statistical Mechanics, Quantum Statistical Physics, Lie Groups and Lie Algebras
- Transfer to MIT after one year of study
Working Experiences
- Extensively leveraging various AI techniques in material and defect engineering, including convolutional neural networks, reinforcement learning methods, generative models, and machine learning force fields (MLFF)
- Co-authored three papers in machine learning-based defect engineering and generative model-powered valence-informed material discovery
- Served as the Seed LLM data partner, led the design of benchmarks comprised of 500+ physics problems, from university to PhD level
- Contributed to ongoing internal projects aimed at improving model reasoning fidelity on symbolic tasks
- Co-led the development of a test-time scaling (TTS) pipeline that enabled gold-medal performance of AI models on IPhO 2025 theoretical problems
- Researched and authored internal tutorials on GPU memory layouts, including CuTe layout and linear layout
- Analyzed the CuTeDSL lowering process to backend kernels using GEMM as a case study, informing optimizations for AI workload deployment
- Authored internal technical notes adopted by the CuTeDSL framework team for performance reference
Achievements
Selected Projects
- Working in a group of four, developed a web platform that converts hand-drawn Mario-style levels into playable platformers in less than 5 minutes by leveraging OpenCV-based image recognition and deployed via Modal
- Won 2nd place in the Modal sponsor track
- Demo: https://demo-description.vercel.app/
- In a team of three, designed and executed a multilingual evaluation of mathematical reasoning on MMATH and GSM8K, comparing frontier APIs (ChatGPT-5.1, Gemini-2.5-Flash, DeepSeek-V3.2) and open-source 8B models across English, Chinese, Spanish, and Thai
- Performed detailed token-length and character-length analyses under five tokenizers to quantify an Encoder Gap between intrinsic and realized representation density, showing up to 5-10% token savings
- Implemented LoRA fine-tuning of Llama-3.1-8B on Chinese GSM8K, improving Chinese accuracy by 8.8 percentage points
- Project repo: GitHub
- Blog: Blog
Skills
- Languages: English (fluent), Chinese (native), Cantonese (basic), Spanish (basic)
- Programming Languages: C/C++, Python, Bluespec, RISC-V Assembly, LaTeX
- Libraries: PyTorch, NumPy, Matplotlib, ASE, PyMatgen, Matformer, MACE, Triton, CUTLASS
Hobbies
- Long-distance running: Completed 1 half marathon with personal best of 1:48:56 (2023 Wuhan Marathon)
- Table tennis
- Movies
- More to explore...
Last updated: December 12, 2025