About Me

• Developing multilingual medical benchmark datasets to evaluate large language models across linguistic and cultural gaps in clinical contexts.
• Contributed to evaluation methodology for medical AI systems, focusing on cross-lingual performance and clinical applicability.

• Conducted computational analysis on single-cell and stem cell datasets, focusing on cell state transitions and perturbation responses.
• Developed models for Virtual Cell Construction, integrating multi-omic data to simulate cellular behavior under various perturbations.
• Contributed to individual-level virtual modeling, aiming to capture inter-donor variability and predict donor-specific cellular responses.
• Performed quality control and computational data processing for mesenchymal stem cell (MSC) datasets, supporting stem cell product characterization.

• Single cell omics analysis
• VQ-VAE / Discrete Representation / Machine unlearning

Predict the area ratio of planets and stars and the confidence level (uncertainty) sigma.

• Achieved the 7th(1%) in the NeurIPS Competition. Achieve my first gold medal.
• We have developed a method that spans from heuristic forward modeling to Gaussian processes, to deep learning, and to gradient boosting trees.
• Become Kaggle Competition Masters.

Reconstruct the direction of neutrinos from the Universe to the South Pole

• Achieved the 21th in the Neutrinos and Astrophysics competition, ranking in the top 3% globally among all participating teams.
• Utilized a 3D point cloud convolution model based on the EdgeConv operator, developed various RNN models, and employed a multi-stage training method grounded in IceCube’s physical principles.
• This achievement also marks my first medal in Kaggle competitions.

Developed a model trained on Electroencephalography (EEG) signals and Spectrogram recorded from critically ill hospital patients to classify a variety of harmful brain activities.

• Achieved 38th place in the Neuroscience and Physiology competition, ranking in the top 2% globally among all participating teams.
• Developed a 1D model based on EEG signals and employed innovative training methods to create an effective 1D+2D multi-modal model.