Used CLIP ViT-B/32 to extract joint image-question features for a real-world, accessibility-focused VQA dataset. Trained various MLP architectures with dropout to handle noisy data and long-tail answer distributions.

Used GoogleNet Inception V3 to extract visual features from TVSum videos for frame-level importance prediction. Trained BiLSTM models on full and downsampled frames to generate summaries using 0/1 Knapsack under a 15% length constraint.

Used Ray to parallelize inference on the IMDb dataset across a cluster using DistilBERT for sentiment classification. Each Ray worker loads the model, processes a batch of texts, and returns predictions, enabling efficient distributed inference.

Implemented binary and multiclass classification using Softmax Regression, SVM, Decision Trees, and Ensemble Methods. Analyzed model performance using accuracy, precision, recall, F1-score, and ROC curves.

Explored Linear, Polynomial, Ridge, Lasso, Elastic Net, and SGD Regression. Applied Gradient Descent methods, Normal Equation, and SVD for optimization.