End-to-end deep learning project for dog breed image classification. Exploratory data analysis (EDA), transfer learning with keras, fine-tuning, and model deployment with flask and docker to Amazon Web Services (AWS) Lambda and Elastic Kubernetes Service (EKS), application deployment to Streamlit.

Credit prediction modeling project using scikit-learn decision tree, KNN, SVM and neural network models to forecast credit default likelihood. Data preparation, EDA, model training and fine-tuning. Deployment with SVM utilizing flask connection and environment containerization with Docker.

Built a GNN architecture using PyTorch Geometric for node classification of the protein-protein interaction (PPI) dataset. Used a combination of Molecular Fingerprint Convolution (MFConv) and Graph Isomorphism Network (GIN) layers, resulting in an F1 score within 5 points of the benchmark while reducing runtime by 67%.

Tested SVM, KNN, naive bayes, logistic regression, and random forest classifiers to predict the presence of fake reviews in the Kaggle Yelp dataset. Used TF-IDF for sentiment analysis and oversampling with SMOTE, noting a 12% increase in recall. Developed an interactive visualization with D3 as an analytical tool to uncover sectors negatively impacted by fake reviews in suburban populations.

Researched the comparative strengths of decision trees, KNN, SVM and neural networks using validation curves, learning curves, wall-clock time and loss curves. Compared tuned performance for the bioresponse dataset, a wide, balanced binary classification problem, to the letter dataset, a long and unbalanced multilabel problem.