Projects

Models trained on easy mazes are tested on much harder ones

When Can Neural Networks Think Longer to Solve Harder Problems?

I stress-tested architectures designed to extrapolate on mazes far harder than their training data, uncovered the shortcut heuristic behind their success, and quantified a data-diversity trade-off. Published at AAAI 2026 (oral presentation).

Read the full story →

The same maze rendered as ASCII text, a pixel array, and a plot

maze-dataset: Open-Source Maze Generation for ML Research

I co-authored a peer-reviewed Python package for generating, solving, and visualizing maze datasets, used in neural-network reasoning research with 80+ GitHub stars. Published in JOSS.

Read the full story →

Satellite imagery of Denver converted to building footprints

Predicting Building Heights from Satellite-Derived Footprints

At Oak Ridge National Laboratory, I trained XGBoost on 898,000 buildings to predict heights from satellite-visible footprints (0.70 m MAE), and showed that adding coarse global height data doesn’t help.

Read the full story →