AI Applied Scientist – PhD Intern, Foundational IQ

About Zillow

Zillow AI’s Foundational IQ group builds the core intelligence that powers search, discovery, and conversational experiences like Zillow Copilot. We combine text, imagery, geospatial and time–series signals to learn rich representations of homes, neighborhoods, and shopper intent, then use them to improve user experiences on Zillow. Our team operates as applied researchers who ship: we prototype quickly, evaluate rigorously (offline metrics and online experiments), and partner closely with product and engineering to bring models to production. We care deeply about helpfulness, safety, and fairness, including domain–specific, fair–housing–aware evaluation, and we invest in modern ML platforms, data governance, and observability to run at national scale. The result is AI that reduces friction, increases confidence, and makes complex home decisions feel simpler for millions of people.

About the role

As a PhD Research Intern on the Foundational IQ team, you will help train and adapt large models that better understand homes and users, advancing representation learning, multimodal modeling, user modeling, and reinforcement/sequential decision–making for real–world problems at Zillow scale. You’ll tailor and evaluate LLMs and multimodal foundation models to our domain, build agentic workflows that plan and act across multi–step tasks, and define success via domain–specific metrics emphasizing helpfulness, safety, and fairness. You’ll move quickly from prototype to impact, running rigorous offline evaluations and online experiments, collaborating with applied scientists, engineers, and product partners, and contributing to platform capabilities that power experiences like Zillow Copilot. Along the way you’ll author clear research docs, share results internally, and have opportunities to publish and present your work.

Key Responsibilities

• Research and develop methods for adapting LLMs and foundation models with Zillow’s domain-specific data
• Build and evaluate multimodal models that combine text, images, geospatial and tabular signals for home and user understanding.
• Explore reinforcement learning and sequential decision-making for long–horizon, user–centric outcomes
• Prototype agentic workflows; define success metrics and run rigorous offline/online evaluations
• Partner across science, engineering, product, and design; share results via docs, presentations, and publications