Rasesh Udayakumar Shetty

Developed a novel architecture using VAEs for detecting AI-generated images with 99% accuracy on similar latent space and 75% on unseen datasets. Finetuned Qwen2-VL model for reasoning.

Technologies: PyTorch, PyTorch Lightning, Gemini API

Designed and trained a custom architecture for GIF Visual Question Answering using PyTorch, incorporating BLIP-2's qformer and Llama3.2-1b.

Technologies: PyTorch, BLIP-2, Llama3.2-1b

Implemented a UNet3+ model for segmenting retinal blood vessels using the CHASE_DB1 dataset. Employed advanced preprocessing techniques, including green channel extraction and morphological transformations, to enhance vessel visibility and suppress background noise. Achieved improved segmentation accuracy and generalization across biomedical images.

Technologies: Python, PyTorch, OpenCV, NumPy

Developed a Generative Adversarial Network (GAN) to generate realistic anime-style faces. The model was trained on a dataset of anime images, utilizing deep learning techniques to capture the unique features and styles characteristic of anime art. The project demonstrates the application of GANs in creative domains, highlighting the potential of AI in art and design.

Technologies: Python, TensorFlow, Keras, NumPy, Matplotlib

Trained Neural Networks to solve Ordinary Differential Equations using PyTorch, gaining insights into function learning using Neural Networks.

Technologies: PyTorch, Neural Networks