Research PublicationsAdvancing Scientific Knowledge

We present IndicBERT, a multilingual BERT model pre-trained on 12 major Indian languages. Our model achieves state-of-the-art performance on multiple downstream tasks including sentiment analysis, named entity recognition, and question answering.

We develop and validate a deep learning system for automated diabetic retinopathy screening that achieves 96.5% sensitivity and 94.2% specificity on a dataset of 100,000 retinal images from Indian patients.

We propose a novel federated learning framework that provides formal differential privacy guarantees while maintaining competitive performance. Our theoretical analysis provides tight convergence bounds under heterogeneous data distributions.

We present a comprehensive analysis of challenges in processing code-mixed text with large language models and propose novel architectures that improve performance by 15-20% across multiple Indian language pairs.

We provide a geometric analysis of optimization landscapes in deep neural networks, proving new results about the connectivity of local minima and proposing improved training algorithms based on these insights.

A comprehensive survey of AI applications for social good in the Indian context, covering healthcare diagnostics, educational technology, and rural development initiatives with measurable impact on millions of lives.