About

I am an AI/ML engineer and applied researcher working on machine learning for climate, health, and society. I did my MS in Computer Science at UC San Diego and I am currently based in India.

Right now I am working on extreme weather prediction, with a focus on tropical cyclones and early warning. Before this, I built LLM systems for healthcare at Oracle, and worked on machine learning for cybersecurity at TCS Research. I also have a background in clinical NLP and AI fairness.

I like finding out whether an idea actually works, then building it into something people use, picking up whatever new method, tool, or field the problem needs along the way.

Selected work

Which Information Matters for Rapid Intensification of Tropical Cyclones? A Global-Basin Feature Ablation Study

Sukhada Ghewari · Shardul Deshpande

NeurIPS 2026 · WiML Workshop poster

Rapid intensification, a jump of 30+ knots in peak winds within 24 hours, is one of the hardest problems in cyclone forecasting and leaves coastal communities little time to prepare. Using 80,000+ observations from 4,000+ storms worldwide (IBTrACS and ERA5, 1980–2025), I tested which information actually drives prediction, using 5-fold cross-validated feature ablations on the training years.

ERA5 environmental predictors add the most: removing them causes the largest drop in AUPRC of any feature group. On storms from 2016–2025, held out from all training and model selection, the final XGBoost model reaches 0.42 AUPRC (about 4× the no-skill baseline) and 0.86 ROC-AUC, and catches about three quarters of rapid-intensification events at a recall-focused threshold.

News