Selected work

Representative Papers

Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems

Feb 18, 2026

This work addresses the challenges of language diversity, document heterogeneity, and deployment efficiency in OCR systems for multilingual and multimodal Indian government documents. The authors propose the Chitrapathak family of methods, which combines end-to-end training of a general-purpose vision-language model with efficient fine-tuning of pretrained OCR models for non-target languages. Additionally, they introduce Parichay, the first dedicated structured information extraction model tailored to nine categories of Indian governmental documents. Experimental results demonstrate that Chitrapathak-2 achieves state-of-the-art performance on Telugu (6.69 character-level ANLS) and ranks second on other languages, while offering 3–6× faster inference. Parichay attains an Exact Match score of 89.8% on key field extraction from government documents, demonstrating both high accuracy and computational efficiency.

0 citationsRead paper

IndicVisionBench: Benchmarking Cultural and Multilingual Understanding in VLMs

Nov 06, 2025

Existing vision-language model (VLM) evaluation benchmarks exhibit Western-centric biases and lack systematic assessment of cultural diversity and multilingual capabilities. Method: We introduce IndicVLM-Bench—the first large-scale, multilingual vision-language benchmark focused on the Indian subcontinent—comprising 5,000 images, 37,000 question-answer pairs across 11 Indian languages, and three core tasks: visual question answering, image–text retrieval, and image captioning. It covers six question types and thirteen culturally grounded themes. We propose cross-lingual alignment and cultural context modeling techniques and release a high-quality parallel annotation corpus. Contribution/Results: Evaluating eight representative VLM families on IndicVLM-Bench, we systematically uncover substantial deficiencies in Indian language comprehension and cultural commonsense reasoning—previously unreported. The benchmark establishes a reproducible, culturally informed evaluation paradigm and provides empirical foundations for developing more inclusive multimodal AI systems.

0 citationsRead paper

Seeing Straight: Document Orientation Detection for Efficient OCR

Nov 06, 2025

This paper addresses the degradation of OCR performance caused by document image orientation misalignment (e.g., due to camera tilt during capture), proposing a lightweight and robust four-way rotation classification method. Methodologically, it introduces an end-to-end fine-tuned pipeline leveraging the Phi-3.5-Vision visual encoder and dynamic cropping. Its key contributions are: (1) the first multi-lingual benchmark—OCR-Rotation-Bench (ORB)—designed specifically for evaluating rotation robustness in OCR, covering English and 11 low-resource Indian languages; (2) a novel architecture optimized for efficient orientation classification; and (3) state-of-the-art results: 96% accuracy on ORB-En and 92% on ORB-Indic, with real-world improvements of +14% for proprietary OCR systems and up to 4× gains for open-source OCR. The code and benchmark are publicly released.

0 citationsRead paper

IndicSuperTokenizer: An Optimized Tokenizer for Indic Multilingual LLMs

Nov 05, 2025

To address the inefficiency of existing subword tokenizers in Indian multilingual large language models—caused by high script diversity and morphological complexity—this paper proposes the first dual-track tokenization framework integrating Byte-Pair Encoding (BPE) and multi-word tokenization, augmented with language-specific pre-tokenization strategies. The framework jointly optimizes fine-grained linguistic modeling and semantic integrity within a unified architecture, significantly improving tokenizer adaptability across diverse Indian languages. Experiments on 22 Indian languages, English, and code data demonstrate an average 39.5% increase in vocabulary generation rate over LLaMA4, a 44% improvement in inference throughput, and competitive performance on multiple downstream benchmarks. This work constitutes the first systematic integration of multi-granularity tokenization mechanisms, establishing a novel, efficient, and scalable tokenization paradigm for resource-heterogeneous multilingual settings.

0 citationsRead paper

VoiceAgentBench: Are Voice Assistants ready for agentic tasks?

Oct 09, 2025

Existing speech assistant benchmarks primarily evaluate isolated capabilities (e.g., ASR or QA), lacking systematic assessment of multilingualism, multi-turn agent interaction, tool invocation, and safety robustness—especially in culturally diverse contexts such as India. This work introduces the first comprehensive agent-oriented benchmark for Speech Large Language Models (Speech LLMs), featuring multilingual mixed-dialogue evaluation, speaker-aware diverse query sampling, adversarial robustness testing, tool-call consistency verification, and safety validation. We innovatively propose a speaker-embedding-based voice query sampling algorithm and construct a high-quality evaluation set comprising over 5,500 synthetic speech queries. Experimental results reveal significant bottlenecks in current Speech LLMs regarding cross-lingual generalization, multi-tool coordination, and defense against adversarial attacks.

0 citationsRead paper

Recent publications

Latest Papers

Designing Production-Scale OCR for India: Multilingual and Domain-Specific Systems

Feb 18, 2026

This work addresses the challenges of language diversity, document heterogeneity, and deployment efficiency in OCR systems for multilingual and multimodal Indian government documents. The authors propose the Chitrapathak family of methods, which combines end-to-end training of a general-purpose vision-language model with efficient fine-tuning of pretrained OCR models for non-target languages. Additionally, they introduce Parichay, the first dedicated structured information extraction model tailored to nine categories of Indian governmental documents. Experimental results demonstrate that Chitrapathak-2 achieves state-of-the-art performance on Telugu (6.69 character-level ANLS) and ranks second on other languages, while offering 3–6× faster inference. Parichay attains an Exact Match score of 89.8% on key field extraction from government documents, demonstrating both high accuracy and computational efficiency.

0 citationsRead paper

IndicVisionBench: Benchmarking Cultural and Multilingual Understanding in VLMs

Nov 06, 2025

Existing vision-language model (VLM) evaluation benchmarks exhibit Western-centric biases and lack systematic assessment of cultural diversity and multilingual capabilities. Method: We introduce IndicVLM-Bench—the first large-scale, multilingual vision-language benchmark focused on the Indian subcontinent—comprising 5,000 images, 37,000 question-answer pairs across 11 Indian languages, and three core tasks: visual question answering, image–text retrieval, and image captioning. It covers six question types and thirteen culturally grounded themes. We propose cross-lingual alignment and cultural context modeling techniques and release a high-quality parallel annotation corpus. Contribution/Results: Evaluating eight representative VLM families on IndicVLM-Bench, we systematically uncover substantial deficiencies in Indian language comprehension and cultural commonsense reasoning—previously unreported. The benchmark establishes a reproducible, culturally informed evaluation paradigm and provides empirical foundations for developing more inclusive multimodal AI systems.

0 citationsRead paper

Seeing Straight: Document Orientation Detection for Efficient OCR

Nov 06, 2025

This paper addresses the degradation of OCR performance caused by document image orientation misalignment (e.g., due to camera tilt during capture), proposing a lightweight and robust four-way rotation classification method. Methodologically, it introduces an end-to-end fine-tuned pipeline leveraging the Phi-3.5-Vision visual encoder and dynamic cropping. Its key contributions are: (1) the first multi-lingual benchmark—OCR-Rotation-Bench (ORB)—designed specifically for evaluating rotation robustness in OCR, covering English and 11 low-resource Indian languages; (2) a novel architecture optimized for efficient orientation classification; and (3) state-of-the-art results: 96% accuracy on ORB-En and 92% on ORB-Indic, with real-world improvements of +14% for proprietary OCR systems and up to 4× gains for open-source OCR. The code and benchmark are publicly released.

0 citationsRead paper

IndicSuperTokenizer: An Optimized Tokenizer for Indic Multilingual LLMs

Nov 05, 2025

To address the inefficiency of existing subword tokenizers in Indian multilingual large language models—caused by high script diversity and morphological complexity—this paper proposes the first dual-track tokenization framework integrating Byte-Pair Encoding (BPE) and multi-word tokenization, augmented with language-specific pre-tokenization strategies. The framework jointly optimizes fine-grained linguistic modeling and semantic integrity within a unified architecture, significantly improving tokenizer adaptability across diverse Indian languages. Experiments on 22 Indian languages, English, and code data demonstrate an average 39.5% increase in vocabulary generation rate over LLaMA4, a 44% improvement in inference throughput, and competitive performance on multiple downstream benchmarks. This work constitutes the first systematic integration of multi-granularity tokenization mechanisms, establishing a novel, efficient, and scalable tokenization paradigm for resource-heterogeneous multilingual settings.

0 citationsRead paper

VoiceAgentBench: Are Voice Assistants ready for agentic tasks?

Oct 09, 2025

Existing speech assistant benchmarks primarily evaluate isolated capabilities (e.g., ASR or QA), lacking systematic assessment of multilingualism, multi-turn agent interaction, tool invocation, and safety robustness—especially in culturally diverse contexts such as India. This work introduces the first comprehensive agent-oriented benchmark for Speech Large Language Models (Speech LLMs), featuring multilingual mixed-dialogue evaluation, speaker-aware diverse query sampling, adversarial robustness testing, tool-call consistency verification, and safety validation. We innovatively propose a speaker-embedding-based voice query sampling algorithm and construct a high-quality evaluation set comprising over 5,500 synthetic speech queries. Experimental results reveal significant bottlenecks in current Speech LLMs regarding cross-lingual generalization, multi-tool coordination, and defense against adversarial attacks.

0 citationsRead paper