Peer-reviewed work, newest first. Citation counts from Google Scholar.
A fully local speech-to-text → language-model → text-to-speech pipeline (Moonshine, Piper, and four quantised LMs from 1.2B to 8B parameters), deployed on an NVIDIA Jetson AGX Orin, a Raspberry Pi 5 and an STM32MP257F-EV1, and characterised on thermostat and washing-machine use cases.
Weights training samples by their distance to the test distribution, down-weighting noisy and out-of-distribution examples. Helps most when labels are few or classes are imbalanced — 2–7 absolute points across benchmarks in the low-label regime.
Extends CART so unlabeled data shapes the split criterion, through sample weights derived from the distance between labeled and unlabeled instances. Grew out of my MSc thesis on optimising XGBoost for semi-supervised datasets.