Explore my technical work across various domains of Machine Learning and Data Science
Innovation: Explored the use of transfer learning and language models for early detection of power outages, achieving significant performance improvements with few-shot learning techniques.
Developed a feed-forward neural network with 12 input variables and Bayesian regularization to predict critical rates. Conducted sensitivity analysis to determine important variables and compared performance with existing correlations.
Built Bayesian optimization and hierarchical time series forecasting models for non-weather-related power outages (NWO), significantly outperforming ARIMA/SARIMA baselines. Defined robust NWOCI metric and leveraged Kats/Prophet techniques.
Built web scraping API for ISONE data collection and developed Robust Quantile Regression Model to analyze hydro effects on electricity price and volatility.
Proposed multi-layer dynamic interconnection model for risk assessment in interdependent systems. Developed data scientific framework using multiresolution data for knowledge discovery across complex systems.
Developed time series regression models including Prophet and Neural Prophet for future horizon prediction. Conducted feature engineering and cross-validation with hyperparameter-tuned Neural Prophet achieving best performance.
Developed embedding and bag-of-words models using Glove representation for sentiment feature extraction. Built ML models to classify review sentiment and compared model performance.
Built multiclass classifier for three-class problem with extensive feature engineering including categorical encoding. Developed three boosted ensemble variations: pre-tuned, hyperparameter-optimized, and class-imbalance-aware models.
Built 11 classification models using NDVI, LST, and Thermal Anomalies for wildfire prediction. Addressed class imbalance with weight balancing and performed comparative analysis with hyperparameter optimization.
Built unsupervised ML model for extractive summarization of UN Sustainable Development Goals articles. Implemented data preprocessing, tokenization, stop word removal, and cosine similarity-based sentence ranking.