Rachakonda Hrithik Sagar · CVIT, IIIT Hyderabad · BharatGen
MS by Research student at CVIT Lab, IIIT Hyderabad and AI Research Engineer at BharatGen. My work spans grounded chain-of-thought reasoning, mask-based document grounding, and large-scale multilingual VLM training — including DoCoG (ECCV 2026), M3Grounder (CVPR 2026), and Patram-7B, India's first foundational document VLM.
DoCoG accepted to ECCV 2026 (Malmö, Sweden) — equal-contribution first author.
M3Grounder accepted to CVPR 2026 main conference — equal-contribution second author.
Merged pull requests in openai/gpt-oss and QwenLM/Qwen3-VL.
Patram-7B released on HuggingFace and IndiaAI AIKosh — India's first foundational document VLM, supporting 22 Indian languages.
Extends grounding beyond answers to the reasoning itself: every step of the chain-of-thought is grounded in textual and graphical document elements through a Grounding Interaction Module with dedicated grounding tokens — so you can see exactly where each reasoning step looks.
Mask-based grounding that handles answers scattered across multiple spans and granularities — words, lines, blocks, figures. Introduces GroundingDocQA: 200K documents, 2M QA pairs. State-of-the-art on DOGR-Bench, BoundingDocs, and MMDocBench.
Supports 22 Indian languages. Built the synthetic data engine that generated 10M QA pairs across 400K documents, ran multi-node training on 64 H100 GPUs, and led all benchmarking for the release.
BharatGen · IndiaAI — Hyderabad
Core contributor to Patram-7B: synthetic data generation at 10M-QA-pair scale, multi-node H100 training, evaluation and benchmarking. vLLM inference optimization with significant latency reduction in production serving.
CVIT Lab, IIIT Hyderabad
Research on mask-based document grounding and grounded multimodal reasoning — leading to M3Grounder (CVPR 2026) and DoCoG (ECCV 2026).
Tech At Core · Hyderabad
Shipped NLP and computer-vision systems for client deployments; engineered a patent-pending drone-ML device combining onboard vision inference with autonomous navigation.
Sharda University — full tuition scholarship for academic excellence · GPA 8.40
openai/gpt-ossMerged pull request — PR linkvLLMProduction inference optimization: significant serving-latency reduction for document VLM workloads 240 × H100Multi-node (30 Nodes) distributed training for Patram-7B pre-training and instruction tuning