• Ph.D. University of Illinois at Urbana-Champaign, Theoretical Physics
• B.S. Beijing Normal University, Physics

Amazon Special Projects

• Spearheaded a cross-functional team of scientists and engineers, delivering a comprehensive protein foundation model evaluation framework within one month; Orchestrated the flawless execution of thousands of AWS SageMaker jobs, ensuring reliability and on‑time delivery of evaluation results in an exceptionally tight schedule

• Designed and maintained a large‑scale LLM training framework built on Megatron‑LM, enabling the ML team to train protein language models up to 30B parameters, achieving MFU> 30 %

• Developed an efficient nd-parallel distributed training framework (DP, TP, FSDP2 with SAC) in native PyTorch and trained a family of protein encoder models where the 1B model achieved MFU > 50% and exceeded the performance of ESM2-3B and ESM2-15B on more than half of BioMap benchmarks

• Applied the PyTorch training framework to Llama models and performed an ablation study that demonstrated the critical role of protein‑sequence diversity in training data

• Innovated a latent‑space diffusion model that generated the top‑performing molecule among millions of candidates, propelling the compound into clinical trials

• Invented a protein design optimization method leveraging fitness model gradients and Hamiltonian dynamics, achieving top-ranking performance across 2024 ML-guided protein engineering cycles — including 95% fitness gain across all 122 targets, 331% increase in average target coverage, and 100% of targets showing statistically significant fitness gains across three distinct design regions, consistently outperforming baseline and competing optimizers

• Engineered a design aggregation and budget allocation algorithm that deduplicated submissions across all design algorithms and assigned fair, balanced budgets to each contributor — reducing runtime from 3 days to under 1 hour while correcting previously unbalanced allocations

• Led ML-to-lab handoff for ML-guided protein engineering cycles, overseeing reverse translation and codon optimization to ensure smooth, successful transitions from computational design to experimental build phase

• Led an internship project culminating in the publication AffinityFlow: Guided Flows for Antibody Affinity Maturation

Clarifai Inc.

• Detection and tracking
• Image embedding
• Fine-grained classification
• Semantic segmentation

• Token classification
• Named Entity Recognition (NER)

• Docker
• Kubernetes and Kubeflow
• CI/CD

Insight Data Science

• Satellite image processing
• Partial convolution inpainting model
• Repo and project page

University of Illinois at Urbana-Champaign

• Exact solution of heat equation in a high-dimensional sphere
• Applied to SVM document classification
• Publication and additional proof

• Effective Dissimilarity Transformation
• Applied to cell line clustering
• Publication

• Reinterpreted spectral clustering through the lens of quantum tunneling
• Publication

• Torch Forward and Backward

  • Native PyTorch distributed training toolkit

• MasonJar

  • Repo-free experiment containerization
  • Build, test, and push Docker images all in Python
  • Reuse experiement images through Python class inheritance

• Mario

  • Kubeflow is very much like Tensorflow (of course they share the last name)
  • Mario is the Kera or PyTorch of Kubeflow
  • Users can intuitively build pipelines without lengthy and confusing API bureaucracy

• Machine Learning Blog

  • Machine Learning Demystified

Theoretical Physics | Mathematics | Probability | Statistical Learning

Protein Engineering | ML-Driven Drug Discovery | Protein Foundation Model

PyTorch | Distributed Training | Diffusion Model | LLM

Classification | Detection | Tracking | ReID | Segmentation

Token Classification | Named Entity Recognition | Transformers

Nvidia Triton | Apple CoreML | Google Protobuf

Docker | Kubeflow | Kubernetes | CI | CD