Lead applied scientist building the ML layer that makes agents feel personal. 14+ years across Zillow and Amazon shipping user-state models, recommender systems, and neural search that decide what an agent should say, retrieve, and do next. Currently leading Zillow's foundational user state action model for agentic experiences, from tokenization and training objectives to offline evaluation and production retrieval.
Core Expertise
User State & Action Models: Sequential transformers over multimodal event streams with multi-task, multi-horizon objectives for next event, attributes, intent, retention, and next-best-action.
Recommender & Search Systems: Two-tower embeddings, ANN retrieval, learning-to-rank, and diversity-aware recommendation for feed and search surfaces.
Agentic Personalization: Dense user representations and predictive heads that ground conversational agents in learned user state, plus simulation-based offline evaluation of agent behavior.
Production Science: Training-data design, offline metrics, calibration, A/B testing, Spark/Databricks pipelines, PyTorch, FAISS, AWS/SageMaker serving.
Experience
ZillowJan 2025 – Present
Applied Scientist, Agentic Foundations · Remote (Northern CA)
- Foundational User State Action Model: Proposed and led a transformer trained from scratch on multimodal home-shopping sequences (search, views, saves, decision-relevant conversation). Defined tokenization, objectives, and representations. Produces dense user embeddings and next-event, attribute, and long-horizon predictions powering intent modeling, personalized retrieval, and next-best-action in Zillow's agentic AI experiences. Blog.
- Agent Simulation & Evaluation: Fine-tuned LLMs to simulate realistic multi-turn shopper conversations; adopted as the offline gate for new agent architectures and major prompt changes.
AmazonJan 2017 – Jan 2025
Applied Scientist, Alexa NewsNov 2023 – Jan 2025
- Personalized News Recommendations: Conceived and launched Alexa News recs in 4 months (POC in 1). End-to-end stack: content scoring, topic clustering, embedding retrieval, and diversity-aware ranking.
- News Summarization: LLM summarization pipeline with headline generation and custom metrics for numerical accuracy and claim validity.
Applied Scientist, FireTV SearchOct 2020 – Nov 2023
- Neural Vector Search: Defined strategy and led a 12-month initiative spanning model, ANN retrieval, and SageMaker serving for plot- and quote-based voice search. Fine-tuned bi-encoders on multimodal catalog metadata over the full video catalog.
- Search Re-Ranker: Two-stage neural re-ranking with training-data design and offline evaluation. +320 bps playback (pointwise), +150 bps further (pairwise).
Research Scientist, Alexa InformationOct 2018 – Oct 2020
- Follow-up Question Recommendation: Co-occurrence, neural CF, and BERT near-duplicate classification over sequential question flows.
- Sports Interest Embeddings: word2vec on follow-action sequences for Sports Briefing and interest picker.
Data Scientist, Amazon Studios · Consultant, AWS ProServeJan 2017 – Oct 2018
- Content & Talent Models: Multimodal plot/actor models for show popularity and bid pricing; entity-level review sentiment for creators.
HomeUnionFeb 2015 – Jan 2017
Data Scientist
- Automated Valuation & Forecasting: GBRT valuation across 110M U.S. homes (4.3% median rent error); published ML home-price index methodology.
SynteractHCRJun 2012 – Feb 2015
Statistical Programmer & Biostatistician
Education
M.S. StatisticsSan Diego State University · 2015
B.S. StatisticsCalifornia Polytechnic State University, San Luis Obispo · 2012
Methods & Technology
PyTorch · Sequential Transformers · Multi-Task / Multi-Horizon Learning · Two-Tower Models · Learning-to-Rank · FAISS · Spark · Databricks · AWS/SageMaker · Docker · Python · SQL