dgsilva

• Built an Informer (Transformer-based) architecture in PyTorch for multiclass state classification across 4 operational states, adapted from the original regression formulation to predict states at 15-minute granularity over a 4-day horizon (384 steps). • Improved the primary business metric (F1-score on the target class) from 30.9% (statistical baseline) to 47.2%, a 53% relative gain, while improving Macro F1 over both statistical and LSTM baselines. • Trained models on Azure Machine Learning GPU clusters using MLflow for experiment tracking across metrics, parameters and artifacts, ensuring reproducibility over hundreds of training runs. • Built the end-to-end data preparation pipeline covering a 7-stage quality filtering process, DBSCAN clustering into 8 behavioural segments, wide-format conversion, and sine/cosine cyclical encoding of temporal features. • Managed the ML lifecycle with the Azure ML SDK, covering compute orchestration, dataset versioning and experiment management, moving from prototyping to standardised, comparable training runs.