Research & Publications

M3DR introduces a comprehensive framework for multilingual multimodal document retrieval, achieving 152% improvement over baselines on cross-lingual retrieval. We release NetraEmbed and ColNetraEmbed, two 4B parameter models supporting 22 languages with state-of-the-art performance across diverse script families including Latin, Devanagari, Dravidian, CJK, and more.

Nayana presents a comprehensive synthetically generated dataset of 3 million document images with hierarchical annotations for training document-centric vision-language models. The dataset spans 22 languages and enables multi-task learning across layout detection, OCR, document retrieval, and more, providing a foundation for universal document understanding.

We introduce Nayana, a scalable framework for adapting Vision-Language Models to low-resource languages using synthetic data generation and parameter-efficient fine-tuning. Using LoRA, we demonstrate effective adaptation across 10 Indic languages (Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) without requiring extensive manually annotated datasets.

A vision-language model specifically optimized for document understanding tasks, capable of processing diverse document formats with high accuracy.

Winner of the 2024 Llama impact grant from Meta, this paper presents a foundation model architecture designed for multilingual and multimodal applications.

A novel framework for autonomous web navigation and task planning using large language models to perform complex multi-step operations.