Wildlife species identification from passive acoustic monitoring recordings in the Pantanal wetlands, Brazil. 234 taxa (birds, insects, amphibians, mammals) classified from Mel-spectrograms via EfficientNet, with multi-round pseudo-labeling on unlabeled soundscapes and OpenVINO inference export.

4th place / 348 teams (private LB 0.939, public 0.964). Wildlife re-identification on 1,895 images across 31 jaguar identities. Final pipeline: EVA-02 Large + DINOv2-Large ensemble trained with ArcFace metric learning, followed by pseudo-label fine-tuning (conf > 0.90), query expansion (k=3), and alpha-QE re-ranking. Backbone search spanned EfficientNetV2, ConvNeXt, and ViT before EVA-02 emerged as best single model (+0.069 LB from 10 → 20 epochs).

4th place / 894 teams, Top 0.5% (private LB F1 = 0.6684). Binary classification of Tidal Disruption Events (TDE) vs. non-TDE nuclear transients from simulated LSST lightcurves with extreme class imbalance (~4.8% TDE). Pipeline: Bazin exponential fit parameters + 80+ hand-engineered photometric features → XGBoost + adversarial validation for train/test distribution alignment.