Loan Default DetectorBuilt a Loan Default Detector to predict credit card applicants at risk of default.Cleaned and analyzed 300,000 application records. Implemented multiple classification models (XGBoost, LightGBM, HistGradientBoosting), addressing class imbalance by integrating SMOTE and undersampling strategies.Orchestrated Hyperopt and MLflow for hyperparameter tuning, achieving a 0.24 PRC-AUC score from the final model, which improved 300% over the baseline.Deployed the model on GCP with Streamlit for internal use, delivering insights for each application prediction.Journal Article Multi-ClassifierBuilt a multi-classifier to classify journal articles with related topics for professors, reducing the working hours of researchers on literature reviews.Applied word embedding techniques (e.g., Word2Vec and FastText) to textual vectorization, and developed a BERT-based model complemented with clustering techniques for topic assignment.Trained and optimized classification models (e.g., Logistic Regression, Random Forest, and XGBoost) and determined XGBoost to be the top-performing model with a 0.55 ROC AUC score and 0.58 Precision score, leading to a 50% reduction in labor costs and an estimated saving of $250,000.

Stream Safeguard: Eutrophication Risk Forecasting for New Jersey's EcosystemsDeveloped a risk forecasting model for New Jersey's streams using historical data on Total Inorganic Nitrogen and Orthophosphate to mitigate eutrophication and enhance ecosystem preservation. Integrated various data sources (spatial, hydrology, water quality) for model training, using classification techniques like Logistic Regression, Random Forest, and XGBoost.Achieved a 58% precision rate with XGBoost, resulting in $1 million in cost savings for lake and watershed restoration, effectively addressing harmful algal blooms.

Seasonal Water Quality Prediction Using Time Series AnalysisBuild a predictive time series model of seasonal water quality of streams to help the New Jersey Department of Environmental Protection effectively make urban and watershed management plans.Collected geological data (land use and land cover) by ArcGIS technique, employed EDA to analysis the feature importance of land use type to water quality. Managed extensive datasets, applying time series analysis to predict water quality with a high degree of accuracy.Achieved a 95% prediction accuracy rate, influencing local water management policies and practices. Visualized data trends and model performance using Matplotlib and Seaborn, facilitating clear communication of results to non-technical stakeholders.House price predictor (water quality related)Build a Hedonic pricing model to predict the property price with water quality dataCollected housing data within watershed management area and select feature among water quality data based on domain knowledge.Train and optimized Hedonic regression models to predict the house price and help the department to save 50% of budget.