GPU-Accelerated 3D UI Layout Optimization

Developed in collaboration with Ben Tatum

As part of my dissertation research on predictive modeling of human performance for 3D user interfaces, I developed a pipeline that uses GPU parallelization to exhaustively evaluate large numbers of candidate 3D UI layouts. The system applies predictive models — including Fitts’ law extensions for 3D interaction tasks — to score layouts based on estimated task completion time across a wide range of user configurations.

The backend uses CUDA to parallelize evaluation across billions of layout candidates, with a Python/FastAPI layer exposing results to a Unity frontend for visualization. DuckDB serves as the analytical backend for querying evaluation results. The goal is to move beyond manual, intuition-driven UI design toward empirically-grounded, model-driven layout optimization for XR interfaces.

Stack: Python, CUDA, FastAPI, DuckDB, Unity, C#