Mohan Raj

A self-supervised Vision Transformer trained with student-teacher EMA distillation on 5,000 unlabelled STL-10 images. Multi-crop augmentation and a centering-and-sharpening loss prevent representation collapse without a single label, reaching 36.5% linear-probe accuracy against 10% chance.

PyTorch · TorchVision · CUDA

Retrieval-augmented chat over biomedical knowledge bases, built on custom embedding and semantic search models tuned for medical text. Improves query relevance by 18% over baseline retrieval and cuts end-to-end latency by 11% through embedding caching and index tuning.

PyTorch · LangChain · Streamlit

A Transformer built from scratch on 1.2M English-German sentence pairs, following “Attention Is All You Need”. Custom attention and positional encoding give 35% faster convergence, reaching 59% BLEU within three epochs at 22% lower memory use.

PyTorch · Transformers

A U-Net segmentation model for medical imaging, trained on the ISIC dermoscopy dataset and evaluated with Dice coefficient metrics.

PyTorch · OpenCV · TorchVision