P-01
KendamaVision
End-to-end computer vision for kendama: a fine-tuned YOLOv8 detector + BoT-SORT tracker, wrapped in a physics-aware post-processor — gravity priors, depth-normalized gates, a soft string constraint — that fills in the fast, blurry tama where the detector misses it, then auto-keyframes a "camera-follows-your-trick" clip. On clips it never trained on (as far as my training records show), the detector finds the ken at about half of hand-marked moments and the tama at under half. A SAM2 auto-labeling → retrain loop turns a few taps per clip into training labels. The writeup has interactive demos running on real tracking data.
P-02
Kendama Ratings
Unofficial Glicko-2 ratings for competitive kendama, rebuilt from Challonge brackets that event organizers published (NAKO, Battle at the Border, EKC and smaller events so far). Every rating is shown as a range, and players are ranked by its low end, so a high place needs a long record behind it. The Python pipeline ingests brackets, reconciles players entered under handles, accounts and bare first names, and tests itself on held-out events: between players it had already seen, the higher-rated one won about three in four matches. A freestyle adjustment, learned from freestyle battles only, orders the Freestyle board.
P-03
ConceptShape — Semantic Geometry Explorer
An interactive 3D "microscope" for the inside of a language model: watch a sentence's tokens move and cluster layer by layer through the residual stream, compare different models' geometry (CKA + Procrustes), and measure the topology of a corpus with persistent homology — a hands-on take on the Platonic Representation Hypothesis. TransformerLens + PyTorch backend, Next.js + React-Three-Fiber front-end.
P-04
SkyWatch
A production Claude-API weather monitor. An hourly Python pipeline ingests Open-Meteo, NWS, and HRRR GRIB2 data into derived meteorological variables; Claude turns them into plain-language briefings, with response caching and cost-stratified model selection across Haiku / Sonnet / Opus.
P-05
biasface
A demographic-bias audit of off-the-shelf face models: a linear probe recovers race from ArcFace and ViT-Base embeddings at ~0.85–0.90 held-out AUC — far above chance, though neither was trained to encode it. Paired with the compliance stakes under GDPR, CCPA, and the NIST AI RMF. Built for EAI 6400 (AI Ethics & Governance).
P-06
Encrypted Vitals
Can a model flag a tumor from a patient's vitals without ever seeing them in the clear? A logistic-regression classifier is compiled to a fully-homomorphic-encryption circuit (Concrete ML / Zama) and runs on encrypted vitals — the server computes on ciphertext it can't read, and the encrypted predictions match the plaintext ones bit-for-bit. Built for EAI 6400.
P-07
GroundTruth
A browser game: classify EuroSAT satellite tiles and race a ResNet18 running live in your browser via ONNX Runtime (WebAssembly) — human vs. AI accuracy, entirely client-side, no server. Built for EAI 6010.
P-08
BoneSim
Dominoes against a Monte Carlo opponent. On its turn, the computer plays out 50 random games for every legal move and picks whichever move won most often. There's no hand-written strategy, just simulation, and it all runs in the browser. Built for fun.
+
More on GitHub
Public repositories, experiments, and works in progress.