Tanish User

At Samsung R&D Institute India in Bengaluru, I built an automatic data sanitization pipeline for generated training datasets using embedding-based zero-shot VLMs like CLIP, plus an OOD scoring function for dataset filtering.

I designed multimodal scoring that fuses OCR text features, image features, and vision-language embeddings into a unified OOD/data-quality decision, and added a statistical thresholding mechanism that works with only in-distribution data when labeled OOD samples aren’t available.

During my PG research at PRISM Research Group, NIT Rourkela, I worked on aerial image object detection in optical satellite imagery under fog, clutter, and occlusion, including tiny object and thermal-aerial detection in multimodal settings.

I also built end-to-end pipelines from my freelance Deep Learning Engineer work: gigapixel whole-slide tumor classification with Multiple Instance Learning, and a cell-level mutation detection pipeline from pathology segmentation masks.