I build human-in-the-loop systems for evaluating, improving, and safeguarding generative models at scale.

Technical lead defining and executing the evaluation strategy for Stitch Fix’s inventory purchasing model, a company-critical system driving millions of dollars in purchasing decisions.

Set priorities for impactful evaluation metrics, built backtesting infrastructure, and surfaced the team’s first daily model performance metrics. Implemented data-backed model improvements to unblock rollout.

Trained computer vision models to predict the performance of new merchandise and productionized the company’s first internal tooling for image-based deep learning.

Investigated methods for improving human–agent collaboration with minimal human data, training hundreds of reinforcement-learning agents and running a human-subject study on Overcooked.

Found that population training fails to improve coordination with humans; led a research team spanning faculty, graduate, and undergraduate collaborators.

Researched the intersection of computer vision and human perception, leading to peer-reviewed publications in top venues (ECCV, CVPR, CHI).

Built a computational model of video memorability, deploying an interactive web platform to collect 10K human judgments and training a model to predict memory decay over time.

Developed novel datasets, interfaces, and models for human visual attention, including the first multi-duration saliency model and a toolbox of interfaces for crowdsourcing attention data.

Advised by Aude Oliva.

Measured and reduced the domain gap between real and simulated data using current techniques in domain adaptation and image-to-image translation.

Integrated neural networks and traditional computer vision techniques into production software