Projects
Research and systems work, each with a short write-up, figures and links.
dKS
A multi-dimensional extension of the Kolmogorov–Smirnov distance that remains a true metric and is invariant to the units each dimension is measured in. Near-linear approximate algorithms make it practical at scale: about 76,000 times faster than the exact baseline at a million samples in 2D. Released as a header-only C++ library with Python bindings.
dKS-3D
The dKS distance in three dimensions. An exact, brute-force-checked baseline sweeps every dominating box and its measured cost grows like N³ — a fraction of a second at n = 400, about 36 minutes per call by n ≈ 8,300. A thinning stage built on a self-balancing walk over dyadic boxes halves the data while every box stays balanced; it is the route to a fast full pipeline.
MotionPI
A privacy-first wearable-sensing platform for an NIH-funded field study of physical activity. BLE wristbands and a smartphone application capture PPG, accelerometry and wristband-computed ENMO, detect moderate-to-vigorous activity on-device, and trigger a survey at the moment it occurs. The deployment sustains roughly 7.7 million records per day with no malformed writes.
Region-aggregated scan statistics
Public-health data usually arrives aggregated to counties or zip codes, and the standard workaround collapses each region to its centroid, discarding its extent. Sampling 20–50 points across each region's actual shape restores the statistical power that collapsing loses, while running about 3,000 times faster than connected-region methods on the continental United States.
Contrast datasets for activation steering
Minimal pairs that isolate a single behavioural axis while holding meaning fixed, used to find steering directions in a model's residual stream and to evaluate how well that control holds. Three sets are public on Hugging Face under CC BY 4.0; the conciseness and positivity sets were built with Martian AI.
Judging by the Cover
A question-blind model that sees only surface features of an answer — length, hedging, punctuation — reaches 68.9% accuracy on TruthfulQA (AUC 0.714), and 14 benchmarks leak similarly. Audit-Prune cleans the benchmark, dropping audit AUC to near chance (0.528) while preserving model rankings (Spearman ρ = 0.915).