Seattle, Washington, United States
As an Applied Scientist with over nine years of experience, I build and deploy machine learning solutions that tackle complex customer problems at scale. My most recent focus has been on recommendation and personalization systems, where I’ve developed and optimized retrieval and ranking models that improve user engagement and experience. I am passionate about experimentation, innovation, and turning data-driven insights into practical, measurable impact for customers. When I’m not working on ML, I’m an avid runner and marathon enthusiast. I am currently training for the Boston and Chicago Marathons in 2026!
Ad platform research
Prime Video Personalization and Discovery • Developed a graph message-passing algorithm that incorporates domain knowledge to learn new item embeddings and categorical features, enabling recommendations for cold-start titles with no prior streaming data (+11.3M annualized hours streamed) and serving as an alternative retrieval source for RAG (+44.4M annualized hours streamed). Presented as an oral talk at Amazon Machine Learning Conference 2025. • Expanded the title-title similarity model from major marketplaces (US, GB, DE, JP) to global regions to enable worldwide model harmonization, addressing challenges including sparse customer signals and varying consumption behaviors across countries. • Built and validated evaluation metrics to measure recommendation quality for customers streaming in secondary languages, bridging data insights with Product Manager goals for international content performance. • Created a persona-based LLM annotation framework that automatically generates high-quality labels for subjective tasks using a persona-selection router to choose optimal annotators from a diverse persona library; applied to search query understanding tasks resulting in an additional 3.67% increase in weighted F1 score, according to distribution of search query occurrences (arXiv:2409.08931). • Designed a mixture-of-retrieval multi-armed bandit algorithm to personalize genre carousel retrievals by balancing users’ long-term subgenre propensities with recent engagement signals, improving offline correlation metrics between recommendations and future customer action evaluation data by 30%. Online A/B testing in progress. • Led weekly science deep-dive sessions to align research directions and foster collaboration among principal applied scientists, principal engineers, and scientists across multiple workstreams
Prime Video Personalization and Discovery • Designed and implemented a modular offline SageMaker pipeline for computing title-title similarities, supporting territory-specific configurations, A/B testing experiments, and rapid ad-hoc experimentation through flexible parameterization and automated deployment. • Established first end-to-end annotations process in Prime Video to obtain human-labeled data for subjective tasks, including establishing a partnership with an internal data services org after evaluating numerous internal and external options, and creating an automated pipeline to generate input manifests for SageMaker Ground Truth jobs and post-process the annotation results.
• Develop deep learning models for image data from manufacturing lines in Gillette • Lead deployment strategy for integrating deep learning into new and existing machine vision systems for real-time inspections in manufacturing
• R&D technology leader for on-line imaging and instrumentation for Baby Care • Deep learning development and integration into on-line imaging inspections • Led development of work process to enable generation of large synthetic image datasets representative of true product images form manufacturing to train deep learning networks
• Managed electrical hardware and PLC software design work for 25 equipment packages, supported integrations at manufacturing sites and global roll-outs • On-line Machine Vision System and Reject System owner for Pampers Preemie Swaddlers product line in Akashi, Japan
In-person and online 1:1 tutoring for high school and university students in computer science, calculus, chemistry, algebra, etc.
Evaluating second-tier automation systems to understand their benefits and risks of implementing them in projects. • Automated a humidifier for the process engineering laboratory using Beckhoff cabinet-less design; utilized IEC 61131-3 standard Structured Text • Created a user-friendly human-machine interface (HMI) using Node.js, JavaScript, and HTML 5 that allows users to interact with Beckhoff controller from a web browser • Performed comparison analysis of Beckhoff automation system to other top-tier automation systems