Competitive ML

EfficientNet-B0 sound event detection combined with Perch v2 embeddings. Macro ROC-AUC ≈ 0.930.

TVT prediction via a LightGBM/XGBoost ensemble with a Bi-LSTM component, rebuilt for fast execution under submission constraints.

GPS and UTC time from rendered scenes, without analytical solar solvers. A three-seed ResNet-34/ResNet-50 ensemble averaged in xyz and sin/cos space.

Predicting pupil correctness from tutoring session transcripts, evaluated on log loss where calibration matters as much as ranking.