DoCoG
DoCoG: Mask-based Multi-Type Grounded Chain-of-Thought for Document QA Oral
European Conference on Computer Vision (ECCV), 2026
Grounded document question answering with step-wise reasoning over textual and graphical elements.
Sai Madhusudan Gunda
Research Engineer, BharatGen · M.S. by Research, Computer Science, IIIT Hyderabad
My research lies at the intersection of computer vision and natural language processing, with a focus on document intelligence: teaching models to read, ground, and reason over complex documents, from modern PDFs to centuries-old palm-leaf manuscripts. I am broadly interested in how vision-language models can move beyond pattern matching toward grounded reasoning, where every prediction traces back to specific visual evidence. This thread runs through M3Grounder (CVPR 2026), DoCoG (ECCV 2026 Oral), and DocLayout-VL (ECCV 2026).
In parallel, I work on OCR for historical Indic manuscripts, where script diversity, physical degradation, and data scarcity collide. UniLipi (ICDAR 2026) unifies OCR across 13 Indic scripts, and CURIO (WACV 2026) tackles curvature and warping in degraded manuscripts. I am currently developing Patram-7B-Instruct at BharatGen, India's National AI Mission, a multilingual vision-language model trained across 256 H100 GPUs.
More broadly, I want to keep asking how models can see, understand, and reason about complex documents in a way that stays faithful to the evidence on the page, especially for the languages, scripts, and archives that today's foundation models still leave behind.
During my M.S., I was advised by Prof. Ravi Kiran Sarvadevabhatla at CVIT.
Email / CV / GitHub / LinkedIn / Google Scholar
News
- May 2026 Two papers (DoCoG and DocLayout-VL) accepted at ECCV 2026, with DoCoG selected for Oral presentation.
- May 2026 Demo paper on Multi-Script OCR for Handwritten Indic Manuscripts accepted at DAS 2026, Vienna.
- May 2026 Converted to full-time Research Engineer at BharatGen after the initial development of Patram-7B-Instruct.
- Mar 2026 M3Grounder accepted at CVPR 2026.
- Feb 2026 UniLipi accepted at ICDAR 2026.
- Nov 2025 CURIO accepted at WACV 2026.
- Jan 2025 Started as Generative AI Research Intern at BharatGen, India's National AI Mission.
Publications
DocLayout
VL
VL
DocLayout-VL: Hierarchical, Open-set, and Promptable Document Layout Segmentation
European Conference on Computer Vision (ECCV), 2026
A promptable foundational document layout model achieving SOTA results across document layout benchmarks.
Indic
OCR
OCR
Multi-Script OCR System for Handwritten Indic Manuscripts Demo
17th IAPR International Workshop on Document Analysis Systems (DAS), Vienna, 2026
A live demonstration of a multi-script OCR system tailored for handwritten Indic manuscripts across multiple scripts.
Uni
Lipi
Lipi
UniLipi: A Unified Multi-Script OCR for Historical Indic Manuscripts
International Conference on Document Analysis and Recognition (ICDAR), 2026
A unified OCR model across 13 Indic scripts with script-aware synthetic data, achieving 6.9% CER and strong cross-script generalization.
M3
Grounder
Grounder
M3Grounder: Mask-Based Multi-Span and Multi-Granular Grounding for Document QA
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026
A mask-based grounding framework that localizes multi-span evidence at phrase, line, and block levels for document question answering.
CURIO
CURIO: Curvature-Aligned and Efficient OCR for Low-Resource Historical Manuscripts
Winter Conference on Applications of Computer Vision (WACV), 2026
A curvature-aware OCR model that handles warped and degraded historical Indic manuscripts, improving recognition accuracy for low-resource scripts.
Experience
Research Engineer
BharatGen · National AI Mission, IIIT Hyderabad
- Converted to full-time after leading the initial development of Patram-7B-Instruct as an intern.
- Continuing to advance multilingual vision-language modeling for document understanding across Indian languages, and scaling training across large H100 clusters.
Generative AI Research Intern
BharatGen · National AI Mission, IIIT Hyderabad
- Developed Patram-7B-Instruct, a multilingual vision-language model for document understanding, OCR, layout parsing, QA, classification, and summarization across Indian languages.
- Built scalable data pipelines for synthetic QA generation, key-value extraction, and summarization for instruction tuning and supervised fine-tuning.
- Implemented efficient multi-node distributed training using DeepSpeed ZeRO-3 on 256 H100 GPUs.
Research Student, Computer Vision
Center for Visual Information Technology (CVIT), IIIT Hyderabad
Advisor: Prof. Ravi Kiran Sarvadevabhatla
- Built ML systems for grounded document QA, historical manuscript OCR, and multimodal document understanding, with a focus on scalable training, evaluation, and inference.
- Published research papers at top computer-vision venues (CVPR, WACV, ICDAR, ECCV), spanning document grounding, OCR, multilingual tasks, and document layout segmentation.
Undergraduate Researcher, NLP
Language Technologies Research Centre (LTRC), IIIT Hyderabad
Advisor: Prof. Radhika Mamidi
- Developed classification models for Indian languages, achieving 92% accuracy on multilingual datasets by fine-tuning.
- Conducted experiments on cross-lingual transfer and low-resource language classification for Indic NLP.
Education
2025 – 2026
International Institute of Information Technology, Hyderabad
M.S. by Research in Computer Science and Engineering · CGPA 9.0/10
2021 – 2025
International Institute of Information Technology, Hyderabad
B.Tech in Computer Science and Engineering · CGPA 8.4/10
Awards & Honors
- Dean's List, IIIT Hyderabad (2021, 2022).
- 1st Place, Megathon Hackathon (IIIT Hyderabad & Qualcomm).
- 7th position, Amazon ML Challenge 2024.
- UGEE All-India Rank 19; JEE Mains 5055; JEE Advanced 3240 AIR.