Models, built from the lens of business.

Forecasting oil spot prices with SOTA transformers (PatchTST, Temporal Fusion Transformer) fed by scraped financial and geopolitical news sentiment scored with Mistral, LLaMA, and BERT — trained end-to-end on GPU HPC clusters.

Prototyped PatchTST against TFT for front-store demand forecasting: patched inputs and channel independence for faster training and stronger performance on seasonal categories.

A weighted loss that emphasizes high-sales weeks and seasonal spikes, so a TFT stops overfitting to the quiet weeks and holds its accuracy through the peaks.

Multi-class genre prediction (10 classes) where neural networks beat tree ensembles on generalizability — with NLP text features and deliberate data-leakage checks so the reported score is real.

A very useful model that tells you whether an image is a hot dog or not a hot dog. A CNN trained from scratch stalled at chance, so a pretrained AlexNet was fine-tuned instead: transfer learning reached 84% in two epochs where the scratch model never beat a coin flip. This is GPU power put to a brilliant cause.