Publications

DoCoG extends document grounding from final answers to the reasoning process itself. Each step of the chain-of-thought is grounded in textual and graphical document elements — tables, figures, charts, and text spans — through a Grounding Interaction Module with dedicated grounding tokens, producing reasoning traces that are verifiable against the source document at every step.

M3Grounder performs mask-based grounding for Document VQA where evidence spans multiple regions and granularities — from single words to lines, blocks, and figures. Alongside the model, we introduce GroundingDocQA, a large-scale dataset of 200K documents and 2M QA pairs. M3Grounder achieves state-of-the-art results on DOGR-Bench, BoundingDocs, and MMDocBench.

Patram-7B is India's first foundational vision-language model for document understanding, supporting 22 Indian languages. Contributions: a synthetic data engine generating 10M QA pairs across 400K documents, multi-node distributed training on 64 H100 GPUs, and end-to-end evaluation and benchmarking for the public release.

Proposes an economically viable solution for women's safety by intercepting the pitch of female voices using sound-frequency sensors — affordable, portable, and deployable in rural areas where CCTV is infeasible. Implemented with sound-frequency sensors, a GSM/GPRS module, Arduino, and Raspberry Pi.

A CNN-based system for early detection of malignant skin cancer, classifying skin lesion images as harmful or harmless to assist dermatologists in early diagnosis of melanoma and other skin cancers.

An AI healthcare chatbot for patient assistance using intent understanding and decision-tree-based follow-up intents (Dialogflow) to predict likely conditions from patient conversations, exploring NLP techniques for conversational medical triage.