My reseach interest are empirical foundations of machine learning: using rigorous controlled experiments to understand the impact of data, model architecture, and other components of ML pipeline to enable efficient and reliable machine learning systems. During my PhD, I applied this to address several fundamental bottlenecks (efficient and scalable training, fine-tuning and inference) for modern-day neural networks (especially large foundational models and graph neural networks). My research accomplishments are evidenced by several impactful publications in top-tier venues, e.g., NeurIPS, ICML, ICLR, ECCV, ICCV, etc.