Projects

Some selected projects showcasing research, implementation, and practical applications.

VisionDesk: Employee activity monitoring using YOLOv8

Python PyTorch Ultralytics YOLOv8 OpenCV NumPy FastAPI RTSP

Mar 2025 – May 2025

  • Developed real-time multi-object detection app using YOLOv8 and FastAPI.
  • Customized dashboard to view employee stats over time.
  • Real-time feed for monitoring employee status.
  • Tuned activity detection with manually annotated workplace footage.
View Code

Vision Transformer (ViT) Custom Implementation

Python PyTorch

Jun 2025

  • Implemented end-to-end Vision Transformer from scratch, including patch embedding and multi-head self-attention.
  • Trained on custom datasets.
  • Visualized attention maps to understand model interpretability.
  • Packaged training, validation, and inference scripts for public use.
View Code

Swin Transformer Variant from Scratch

Python PyTorch Swin Transformer components NumPy

Jun 2025

  • Built window-based multi-head self-attention with shifting mechanism per Swin Transformer design.
  • Demonstrated performance gains in classification tasks compared to standard ViT.
  • Automated model checkpointing and fine-tuning utilities.
  • Conducted experiments on patch/window size sensitivity.
View Code

CutMix & MixUp Implementation in PyTorch: Regularization for Model Training

Python PyTorch torchvision

Jun 2025

  • Implemented CutMix and MixUp data augmentation strategies to regularize CNN training.
  • Integrated both techniques into PyTorch training pipelines with custom datasets.
  • Compared model generalization and convergence behavior under different augmentation regimes.
  • Visualized mixed inputs and label distributions for interpretability and debugging.
View Code

Knowledge Distillation: ResNet50 → MobileNetV3 on STL‑10

Python PyTorch ResNet50 teacher MobileNetV3 student

Jun 2025

  • Trained teacher-student pipeline transferring knowledge from a high-capacity ResNet50 to efficient MobileNetV3.
  • Tuned temperature and loss weighting to balance mimicry and ground truth.
View Code

SimCLR Contrastive Learning Framework

Python PyTorch torchvision contrastive loss

Apr 2025

  • Implemented SimCLR pipeline with dual-view augmentations and projection heads.
  • Pre-trained encoder on unlabeled image dataset achieving robust representation scores.
  • Evaluated learned embeddings on downstream tasks.
View Code

Semantic Segmentation Pipeline

PyTorch U-Net variants

Dec 2024

  • Implemented U-Net semantic segmentation architecture for varied datasets.
  • Implemented Vanilla U-Net and U-Net with pre-trained Resnet18 encoder.
  • Visualized predictions versus ground truths for qualitative validation.
  • Modularized code for dataset swapping and transfer learning.
View Code

Conditional GAN in Keras: Class‑Conditioned Image Generation

Python TensorFlow/Keras NumPy Matplotlib

Nov 2024

  • Built a Conditional GAN using Keras to generate MNIST digits conditioned on class labels.
  • Designed generator and discriminator networks with label embedding and concatenation.
  • Trained the model to produce class-consistent samples with minimal mode collapse.
  • Visualized generated samples across epochs to monitor training quality.
View Code

Flower-Classification-Using-NN: Deep neural networks for fine‑grained image classification

Python Keras TensorFlow NumPy Pandas Jupyter Notebook

Aug 2023

  • Built and trained multiple CNN architectures to classify flower species from image datasets.
  • Employed techniques like transfer learning and data augmentation to boost classification performance.
  • Evaluated models using metrics like precision, recall, and top‑k accuracy.
View Code

Flood-Segmentation: Flood image segmentation model(s)

Python TensorFlow Keras OpenCV Jupyter Notebook

Jun 2023

  • Designed and implemented a U-Net-based segmentation pipeline to delineate flood-affected regions from satellite imagery.
  • Optimized training and data augmentation strategies to improve segmentation accuracy.
View Code