• Built an internal agent-driven experiment analysis system that automatically queries, ranks, and summarizes historical launch proposals, cutting manual analysis time and enabling faster, more consistent decision-making across teams

• Designed and launched a client-to-server ad invalidation framework that filters already-invalid ads (e.g., app already installed) early in retrieval, delivering +0.1%-0.2% incremental Facebook ads revenue at scale

• Led multiple ads retrieval infrastructure optimizations - including inactive ad filtering and storage redesign, that reduced CPU and storage usage by up to 98%, saving roughly $800K in annual compute cost while improving system efficiency

• Led YouTube Android client-side video playback and download efforts across YouTube platforms to improve playback latency, app system health, video download reliability, and engagement

• Launched Android streaming response caching strategy to improve 22.5pp fast Shorts livestream transition rate, and contributed to 11.5% increase to Shorts livestream playbacks

• Recognized as Google Perfy Gold Awardees (66 total Googlewide) in 2024Q3 aimed for exemplar methodology/result on high-value performance and efficiency issues (latency, compute, memory, spindles)

YouTube:

• Privacy Compliance: Designed and deployed client-side stream directory migration to fix security vulnerability for over 2 billion users

• Download Efficiency: Addressed and resolved critical issues leading to stalled video downloads, achieving an 83% reduction in video downloads stuck over 3 minutes by fine-tuning retry strategies

• User Experience: Reduced video download cancellations by 20% by resolving a multi-device casting race condition

• App Stability: Enhanced YouTube Music Android app stability, reducing total crashes by 15.6% through channel thumbnail download via https

• Database-Cache Consistency: Improved download performance with a 2.2% reduction in download failures by ensuring video download database-cache consistency

• Database IO Efficiency: Reduced user ANR (app not responding) events by 5% by eliminating redundant data encryption and decryption processes during database operations

• Data Integrity: Implemented a clean-up strategy that led to a 49% decrease in corrupt downloaded streams

Google Research (20% project):

• Collaborated with interdisciplinary teams of computational chemists, statistician, and software developers from research institutes and AI-driven drug discovery company to develop novel active learning algorithm and improve performance in small molecule hit-to-lead conversion flow

SimuΧ– A Computer Model to Simulate Soft Bio-hydrogel Material 2014 – 2018

Molecular Computational Lab, University of Maryland, College Park

• Proposed a state-of-art computer model for biomaterial used in artificial skin synthesis to reduce the computational cost for simulations by 20 times

• Developed analytical applications to identify crosslinking patterns of chitosan polymer network using clustering algorithm implemented in Python to reveal 3D molecular structural properties

• Collaborated with BioChip groups to characterize the hydrogel deformation behavior based on 2TB simulation dataset and improve the understanding of biofilm fabrication

WEPPRO - A Highly Efficient Peptide Model with Optimized Drude Dipoles 2013 – 2018

Molecular Computational Lab, University of Maryland, College Park

• Modeled Water-Explicit Polarizable Proteins (WEPPRO) to obtain unbiased protein structures, a critical prerequisite for the design of drugs targeting the Alzheimer’s disease

• Simulated four million frames of molecular trajectories and unravel the underlying molecular mechanisms for self-assembly processes via statistical algorithms implemented in Python

• Created molecular images featured in peer-reviewed scientific journal cover using ray tracing system implemented in Visual Molecular Dynamics (VMD)

WEPMEM – An Optimized Lipid Model to Simulate Bilayer Dynamics 2012 – 2015

Molecular Computational Lab, University of Maryland, College Park

• Developed Water-Explicit Polarizable Membrane Model (WEPMEM) for lipids and cholesterol related to coronary artery disease for highly efficient computational simulations

• Investigated the lipid aggregation using machine learning algorithms implemented in Python scikit-learn and automated the analysis of simulation dataset

• Optimized a job management system to automate WEPMEM simulation setups in high-performance computing platforms using bash/csh

• Investigated large-scale genome data to gain insights of type-II diabetes using BLAST, Clusteralx2, and Python (BioPython) and automated the data retrieval process from genome database using Python and SQL

• Automated the process to retrieve data from genome database and improved the efficiency of data collection using Python scripts with provided API