You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I make LLMs more interpretable with activation steering, prove things about high-dimensional statistics, and build the systems that put both into the field.
Steering pipelines with layer/α sweeps on Llama 3.1 and Qwen-2.5, LLM-judge + JSON-schema eval guardrails, benchmark auditing — shipped three contrast datasets with Martian AI; earlier, a dual-stream CNN with attention for hyperspectral change detection.
⚙️ Systems, infra & wearables
Sole developer of MotionPI, a privacy-first wearable-sensing platform behind an NIH field study: wristband → Flutter app → REST API → MongoDB, offline-first schema-validated sync at ~7.7M records/day with zero malformed writes; containerized services and reproducible builds across the stack.
🧮 Algorithms
dKS: header-only C++ core with pybind11 Python bindings, CMake/Make builds, correctness tests against an O(n³) reference — plus sampling-based spatial scan statistics, ~3,000× faster than FlexScan.
Reported and reproduced a Flutter SDK bug (flutter#166937) — confirmed and P2-triaged by the Google framework team.
Header-only C++/Python library for the multi-dimensional Kolmogorov–Smirnov distance — near-linear algorithms, ~76,000× faster than baseline at 1M samples
Code and figures for 'Sampling for Region-Aggregated Spatial Scan Statistics' — higher detection power, ~3,000× faster than FlexScan (arXiv:2607.01451)