Hi, I'm Kabir J.

I'm a Research Intern at Microsoft Research Bangalore (from November 2026), working on embeddings, and a Data Science student at IIT Madras.

Most of my work comes back to one question: what is actually going on inside a model, both when it works and when it doesn't? At Precog, IIIT Hyderabad, with Prof. Ponnurangam Kumaraguru, I studied interpretability in multimodal LLMs, looking at how grounded conversational models connect what they say to what they see. Before that, at VISTA Lab, IISc Bangalore with Prof. Punit Rathore, I worked on self-supervised learning for 3D medical imaging, which became our IJCAI-ECAI 2026 paper. I got to present it in Bremen in August.

I learn best by building things from scratch. I've re-implemented 18+ deep learning papers in NumPy and PyTorch, trained a small Mixtral-style mixture of experts, and wandered into side quests like getting LLMs to be funny on purpose. That same habit is why there's a neural network training on this page.

Outside research I'm usually reading, playing the blues on my electric guitar, or watching shows and anime. I'm always happy to talk about interpretability, representation learning, or anything else, so feel free to reach out.