Principal Applied Scientist (User Modeling)

Zillow

3d•$185,800 - $312,400•Remote

About The Position

The Principal Applied Scientist will join Zillow’s Agentic Foundations team, a group of 15–20 applied scientists in the broader Agentic AI organization, to develop foundational AI technologies for home-shopping experiences. The role focuses on user modeling: building models that help agents understand shoppers’ journeys, goals, and how interventions affect subsequent outcomes. The scientist will address complex and ambiguous problems end-to-end, partnering with product, engineering, science, platform, and data teams to bring advances in sequence modeling, causal inference, and reinforcement learning into production. Zillow is building agentic capabilities across the home-shopping journey, including search, personalized guidance, offer strategy, and financing, at a scale of hundreds of millions of requests per day. The role owns end-to-end LLM post-training pipelines, including supervised fine-tuning, direct preference optimization, and reinforcement fine-tuning, running training on GPU infrastructure. It also develops multi-category reward models, assesses live agent trajectories using LLM-as-a-Judge frameworks, and translates evaluation rubrics into reward functions that generalize to production traces. The position provides technical leadership and mentorship while translating state-of-the-art AI research into production impact. This is a remote position. Employees do not have a permanent corporate office workplace and must identify their physical work location to the Company. U.S. employees may live in any of the 50 states, with limited exceptions. The position is eligible for equity awards; salary ranges vary by location.

Requirements

PhD in Computer Science, Statistics, Economics, Machine Learning, or a related quantitative field, with depth in reinforcement learning or causal inference.
Hands-on experience modeling user behavior at scale, including sequence models over event streams, sessions, or trajectories, and predicting user intent or long-term outcomes.
Practical depth in causal inference—such as propensity scoring, treatment-effect estimation, or off-policy evaluation—or reinforcement learning applied to real sequential decision problems.
Ability to independently scope and solve ambiguous problems end-to-end while providing technical leadership to scientists and MLEs.
Working proficiency in Python and ML frameworks such as PyTorch or TensorFlow, sufficient to run experiments and read, review, and debug code.

Nice To Haves

Background in economics or econometrics, such as demand modeling or consumer choice.
Published work in reinforcement learning, causal inference, survival analysis, or user modeling/simulation.
Experience with generative sequence models or conditional language generation for behavioral or conversational simulation.
Experience with rare-event modeling and probability calibration at scale.
Familiarity with two-sided marketplaces, real estate, or long-horizon consumer journeys.

Responsibilities

Own LLM post-training pipelines, including SFT, DPO, and reinforcement fine-tuning (RFT/GRPO), and run training end-to-end on GPU infrastructure.
Build and train fast, multi-category reward models (PRMs) that emit scalar signals from contrastive preference pairs.
Design on-policy and online assessment by scoring live agent trajectories using LLM-as-a-Judge frameworks.
Translate offline evaluation rubrics into generalizable reward functions that work on arbitrary production traces.
Independently drive complex work end-to-end in collaboration with partner teams.
Provide technical leadership and mentorship, and help translate state-of-the-art AI research into production impact.
Build user world models that help agents understand shoppers’ journey stages, goals, and how actions change outcomes.