Projects
Selected research and systems work — each is a short case study with figures, numbers, and links.
Judging by the Cover
LLM evaluation · Benchmark auditing
Could you ace a truth quiz without reading the questions? On TruthfulQA, yes — six style cues pick the "true" answer at AUC 0.714. Our Audit-Prune cleanup drops that to chance (0.528) while preserving model rankings.
Full write-up & paper — coming soon
Statistics · Algorithms
Why should comparing two datasets change just because you measured in grams, not kilograms? It shouldn't — dKS is a unit-invariant true metric that stays fixed under rescaling and runs ~76,000× faster than the baseline at a million samples.
Wearable biosignals · Free-living health data
A wearable health model is only as good as its data — and free-living signals are where quality breaks. MotionPI captures clean PPG (HR/HRV), accelerometry, and ENMO with real-time on-device MVPA detection, sustaining ~7.7M records/day with zero malformed writes for an NIH study.
Spatial statistics · Python
Where do you plot a disease case reported only as a county-wide total? Instead of one dot at the county's center, we scatter 20–50 points across its real shape — recovering outbreaks the shortcut misses, ~3,000× faster than region-based methods.