Developing and testing new deep learning architectures for subseasonal-to-seasonal (2- to 6-week) weather forecasting. Experiments include variational autoencoders for weather data, treating the resulting latent representations the way language models treat tokens, and extending context using retrieval-augmented generation — adapting methods popular in language modeling to push the limits of AI-based weather forecasting.
Building a conditional diffusion model to predict the next atmospheric state from the two preceding weather states. The model iteratively refines a noisy candidate state through a denoising network based on an encoder-processor-decoder architecture, operating on a refined icosahedral mesh for global weather forecasting.
Investigating whether preference-based fine-tuning improves the naturalness of English-to-Hinglish translation compared to standard supervised fine-tuning alone. Involves building a large parallel corpus (around 350,000 sentence pairs) from real science lecture transcripts, fine-tuning a translation model on it, and refining it using human judgments of translation quality — aiming to make English-medium STEM lectures more accessible in natural, Roman-script Hinglish while preserving technical terms.
The boreal summer intraseasonal oscillation (BSISO) — a wave drifting north over India — controls monsoon active and break spells by modulating the low-pressure systems that deliver the rain. A transformer model forecasts the BSISO two weeks ahead; that forecast, combined with observed low-pressure-system labels, feeds a logistic model giving active and break spell probabilities.
Building a large language model-based framework for scientific literature review of Northeast Monsoon research, in two parts: extracting and summarizing key information from a large collection of research papers, and then verifying the accuracy, robustness, and reliability of the generated outputs — developing methods to assess LLM-generated scientific summaries without manually checking every paper.
Using the reduced-dimensional latent representation of global weather states to cluster the latent space and construct a dictionary of representative weather states. Analyses the clusters to determine the probabilistic relationship between the current weather state and its subsequent state.
Developing a diagnostic framework to study how different capabilities — reasoning, vision, long-context processing, verification, and action — interact within multi-agent AI systems for scientific problem-solving. Focused on how agents communicate with one another, comparing strategies from sharing only conclusions to sharing full reasoning and uncertainty, and how this affects downstream decision-making and error propagation.