Relevant Coursework: High Performance Machine Learning, Theoretical Foundations for LLMs, High Dimensional Stats for Biomedical Data, Advanced Programming, Data Structure, Calculus III, Linear Algebra, Machine Learning - Stanford (Coursera), Deep Learning Specialization - DeepLearning.AI (Coursera)

• Developing a multimodal transformer model to predict binding affinities of protein-molecule bindings.
• Optimized the training and inference pipeline of protein language models by up to 30% by reimplementing models using flash attention and provided a validation method to evaluate data from wet-lab experiments.

• Designing algorithms to select a maximally diversified set of proteins for molecular dynamics simulations, generating data to support subsequent model development for predicting protein conformation trajectories.
• Developing methods to extract protein conformational ensembles from AlphaFold2 through latent space exploration and systematically created a benchmark library to compare with existing methods.

• Established fine-tuning pipelines with Low-Rank Adaptation (LoRA) and Supervised Fine-Tuning (SFT) for four pre-trained DNA language models to evaluate its representational power for regulatory genomics.
• Developed and implemented evolution-inspired data augmentations (EvoAug-TF) in TensorFlow for genomic deep neural networks and demonstrated its improvement in generalization and interpretability.
• Designed and evaluated more than 100 deep-learning models for predicting DNA promoters' expression rates using Python, TensorFlow, and WandB in the 2022 DREAM Challenge. Placed 7th in the final leaderboard.

• Developed a continuous individual crisis aid alert system (CICaidA), manufactured the prototypes, and tested the hardware backend interface system.

• Guided students in ESE123 to conduct experiments and write lab reports in the lab; hosted weekly office hours.

• Developed deep learning models to predict small molecule-protein interactions using the Big Encoded Library for Chemical Assessment. Implemented over 40 types of DL models, including CNNs, GNNs, Transformers, RNNs, and GDBT Models, and finally developed a robust solution after using up all 480 submissions.

• Facilitating mentor connections and providing programmatic support for cohorts of 6-8 early-stage startups during Columbia's intensive 8-week Almaworks accelerator, helping teams prepare for the culminating investor pitch Demo Day to drive fundraising success.

• Hosted biweekly machine learning/python hands-on workshops to cultivate members in their interests in AI.
• Collaborated with various professors and graduate students to provide research opportunities to undergraduates.