Vishnu Kadam

Vishnu Kadam

BS-MS · IISER Pune

I'm a BS-MS graduate from the Indian Institute of Science Education and Research, Pune. My interests lie in all things deep learning — from theory to the latest applications. Recently I've been drawn to Mechanistic Interpretability and am working my way through it seriously.

Now

Investigating mechanistic interpretability techniques on AI weather models — trying to uncover what these models actually learn internally.

Interests

Mechanistic Interpretability Deep Learning Weather Prediction NLP

Publication

CurvePotGCN – A graph neural network to predict protein–protein interactions using surface curvature and electrostatic potential as node-features

2026 PDF ↗

Projects

Weather Prediction · Master's Thesis

Started as a curriculum project and grew into my thesis. I trained and fine-tuned an SFNO-based weather prediction model, then benchmarked its regional forecasting capabilities after fine-tuning. Currently extending this with mechanistic interpretability to understand what the model has actually learned.

Protein–Protein Interaction

My first research project was deep learning for protein–protein interaction under Dr. Arnab Goswami. I handled everything end-to-end — dataset creation, model architecture decisions, training, and benchmarking. It was my first real exposure to what research actually demands, and it led to my first publication.

Diffusion Models

An in-depth dive into score-based diffusion — I implemented annealed Langevin dynamics from scratch and built extensive visualisations to develop an intuition for the process.