San Francisco Bay Area
Senior AI Researcher at Chan Zuckerberg Biohub working at the intersection of AI for Science, multimodal foundation models, and large-scale scientific learning. I design efficient training pipelines and scalable architectures for protein and genomic predictions. Previously, I’ve worked across academia, national labs, and industry, at Lawrence Livermore National Laboratory, the University of Washington, IBM Research, Amazon, and Siemens, leading and contributing to projects spanning drug discovery, medical time-series modeling, EHR analysis, and medical QA systems. My work emphasizes translating advanced ML methods into robust, real-world scientific and biomedical applications.
AI for Biology
Visiting Researcher for DOE projects.
Worked on several DOE-funded projects at the intersection of time-series modeling, health data, and foundation models, driving innovations in predictive analytics and AI for healthcare. • Foundation models on bio-sequences and multi-million patient health records, given by Kaiser Permanente, to forecast temporal patterns and early disease diagnosis. • AI agents to detect defects in 3D-printed structures using multi angle image analysis and multi-variate time-series data from manufacturing processes. • Optimization of large-scale foundation models (de novo) national lab GPU clusters such as Tuolumne cluster (among the most powerful supercomputers globally).
Contributed to DOE-funded projects on time-series modeling and health data, developing AI methods for predictive healthcare analytics.
Researcher under Prof. Tim Althoff (Behavioral Data Science Group), focusing on LLMs capable of understanding and reasoning with multi-modal time-series data, including both images and text. • Worked on applying LLMs to time-series forecasting, identifying limitations in existing models and proposing simpler, scalable architectures. Awarded NeurIPS 2024 Spotlight and featured in multiple articles, including IBM blogs, referred as “One of the biggest revelations of 2024”. • Released a benchmark time-series understanding dataset, adopted by companies like Apple and ByteDance to evaluate LLM reasoning over temporal and textual data. • Developed a LLM prompt-injection dataset (7K+ monthly downloads on HuggingFace) and a defense framework treating prompts as programs, influencing GitHub’s test-case generator.
With the WatsonX Data & AI team, I worked on enabling Watson-Core to perform business intelligence tasks such as data denoising, feature aggregation, etc., using only text commands over IBM cloud. Developed the first functional prototype replacing Watson-Core with an LLM-powered backend for business tasks and led a tutorial on automated intelligence agents at a major NLP event.
Worked with the ML team on time-sensitive reward distribution for users in Amazon Pay. Analyzed payment records to model user-item preferences and predict future purchases for personalized and well-timed coupon recommendations, and tested methods to handle missing events.