projects

2025

LLM from Scratch
Implemented BPE tokenizer, Transformer decoder architecture, and autoregressive pretraining pipeline end-to-end in PyTorch
code

Vision Transformer from Scratch
Implemented ViT architecture and training pipeline from scratch in PyTorch for image classification
code

2024

Semantic Segmentation
Built U-Net for urban scene understanding; achieved 88.5% test accuracy and 0.40 mean IoU
code

GAN for Image Generation
Implemented Generator and Discriminator from scratch for MNIST image synthesis
code

Human Activity Recognition
Benchmarked LSTM, BiLSTM, and 1D-CNN on multivariate time series; best F1 of 0.630 with 1D-CNN
code

Multivariate Time Series Forecasting
Two-layer BiLSTM on 7-feature sequential dataset; RMSE of 4.03
code

Clustering Benchmarking
Compared LDA, QDA, KNN, and SVM; best F1 0.800 with SVM. K-Means clustering with PCA on MNIST achieving 0.790 purity
code

Boston Housing Price Prediction
Built a feedforward neural network for regression; achieved very low MSE of 0.00457
code

Wine Quality Prediction
Developed neural network with regularization (dropout, early stopping); achieved MSE of 0.0127
code

Heart Disease Classification Benchmark
Compared LDA, QDA, and KNN models; achieved best accuracy of 0.7527
code

DILLEMA: Robustness Testing for Vision Models
Designed framework to uncover failure cases in deep vision systems; detected up to 53.3% misclassification in ImageNet models
code

Statistical Learning: Linear Regression
Implemented OLS regression with hypothesis testing; achieved R² score of 0.897
code

Iris Classification with SVM
Applied SVM and benchmarked against classical models; achieved F1 score of 0.7999
code

Leaf Image Classification
Trained deep CNN (Inception-ResNet) for 15-class classification; achieved PRC score of 0.8867
code

CNN for Image Classification (MNIST & CIFAR-10)
Implemented LeNet and VGG-like models; achieved F1 0.9899 (MNIST) and 0.7866 (CIFAR-10)
code

Transfer Learning with MobileNetV2
Improved CNN performance via fine-tuning vs retraining on limited datasets
code

Bias-Variance Trade-off Analysis
Analyzed impact of model complexity on generalization in regression tasks
code

Image Augmentation for CNNs
Improved model robustness using augmentation techniques (rotation, flip, brightness)
code

Feature Selection for MPG Dataset
Applied forward/backward selection with cross-validation; achieved R² of 0.8462
code

Fully Convolutional Network for Heatmaps
Built FCN to generate spatial heatmaps for object localization without dense layers
code