George Hu

A prior simulation, rebuilt on real model weights. I attach DeeBERT-style early-exit heads to a frozen BERT-base encoder and route by softmax-entropy confidence, measuring real energy with powermetrics on Apple Silicon. Result: ~37% fewer layers and ~37% measured energy at 98.8% of baseline accuracy — with no statistically significant accuracy loss across 5 seeds. Independent research; code on GitHub.

Three minutes to pick your FTC alliance. Raw scores lie — OPR via Gaussian elimination separates signal from noise. Bayesian shrinkage, 2,000 Monte Carlo sims per pairing, synergy scoring. Pulls live data from the FTCScout GraphQL API. Bilingual (EN/ZH), auto season detection. Built by Team IK19859. Used at real competitions. Live at scoutselect.org.

AI rowing technique analysis. A phone video becomes pose keypoints, segmented strokes, and nine biomechanical checks per stroke with traffic-light coaching cues. Crew mode scores a whole eight seat by seat. Camera-angle correction, honest confidence intervals, club portal, bilingual throughout, plus a native WeChat Mini Program. The whole product, video processing included, runs on one modest server. Live at rowinglab.org.

A five-model stacked ensemble — RandomForest, ExtraTrees, LightGBM, HistGradientBoosting, GradientBoosting blended via a Ridge meta-learner — over ~75 temporally-safe engineered features, causal ELO ratings, and a vectorised Monte Carlo engine running 10k–50k simulated races a second. Walk-forward backtested across four F1 regulation eras, 2014–2026. Rich CLI + Streamlit dashboard sharing one analysis core.

An AI plugin suite for FTC teams that refuses to guess. Five skills — team config, code generation, hardware lookup, rule check and code review — each grounded in real sources: verified rule citations, structured hardware specs, and library documentation. Config-gated, so it will not write robot code before the robot's configuration is confirmed. Python and Java over MCP; works in Claude Code, VS Code, Copilot CLI, Cursor and Gemini CLI.

A behaviour-tree decision engine for FTC: given the state of the match, the robot and the field, which behaviour should run right now, and how does it get preempted the instant that stops being true. Zero heap allocation on the steady-state tick path, so no GC jitter mid-match. Pure-Java core with no Android or FTC-SDK dependency, unit-testable off-robot, plus an FTC bridge for FTCLib and SolversLib.

A modernised command-based stack for FTC Team 19859, rebuilt on current SDK versions. SolversLib for control, Pedro Pathing for autonomous navigation, Panels telemetry, goBILDA Pinpoint localisation, Limelight3A vision. Commands, subsystems, hardware abstraction and opmodes are kept as separate layers, with ADRs, a wiring map, a six-phase bring-up checklist and GitHub Actions build validation.