May 2025 – Jul 2025 · 2 mos · Gandhinagar, Gujarat, India · On-site

• Engineered and fine-tuned Vision Transformer based medical imaging pipelines for lung lesion segmentation on chest X-rays using custom hybrid Focal+Dice loss, optimizing model convergence and interpretability on clinical datasets using IPR's HPC cluster.
• Developed scalable preprocessing and automated mask-generation pipelines from JSON annotations, streamlining end-to-end experi mentation, metric visualization, and reproducible training workflows across CNN, ViT, and Transformer architectures.
• Researched multimodal AI integration using T5/BART transformers for radiology report generation and explainable medical AI; optimized T5-based pipelines to achieve over 2x lower evaluation loss than BART baselines during experimentation.
• Exposure: Python, PyTorch, TensorFlow, Linux, Cluster Computing, HuggingFace Transformers, Vision Transformers, Explainable AI (XAI), Medical Imaging, Multimodal Learning.