Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{203941,
        author = {Ritu Gaur and Mr. Gaurav Kumar and Ms. Archana Jain and Mr. Aditya Yadav and Ms. Rashmi Singh and Dr. Deepak Kumar Gupta},
        title = {AI-Powered Intelligent Document Processing Using LLMs, Deep Learning, and Knowledge Graphs},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {2894-2905},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=203941},
        abstract = {Intelligent Document Processing (IDP) has become a critical enabler of digital transformation in the Banking, Financial Services, and Insurance (BFSI) sector, where institutions must extract reliable structured data from large volumes of heterogeneous, semi-structured, and unstructured documents under stringent regulatory constraints. Traditional Optical Character Recognition (OCR) and template-based pipelines are brittle to layout variation, struggle with noisy scans and handwriting, and provide limited semantic understanding, which leads to high exception rates and costly manual review. Recent advances in deep learning–based computer vision, Transformer architectures such as BERT and Layout LM, and large language models (LLMs) have significantly improved document understanding, but LLMs remain susceptible to hallucination and inconsistent field-level accuracy in high-stakes domains. This paper proposes a hybrid IDP architecture that tightly integrates (i) a CNNbased image pre-processing and OCR enhancement layer, (ii) an LLM-driven semantic extraction layer with task-specific prompt engineering, and (iii) a domain-specific financial knowledge graph implemented over a property-graph store with RDF-compatible ingestion for constraint-based validation and entity reconciliation. The system targets key BFSI use cases such as loan applications, KYC documents, bank statements, and insurance claims, including multilingual and handwritten content. Experimental evaluation on a multi-document BFSI corpus demonstrates that the proposed architecture improves field-level F1-score by 6–10 percentage points over strong CNN+ Transformer baselines, while reducing hallucination-induced validation errors by more than half through graph-based consistency checks. Detailed analyses of precision, recall, latency, and scalability indicate that graph augmented LLM extraction can achieve near real-time straight through processing rates suitable for production-scale financial workflows. The findings highlight the effectiveness of combining LLMs, deep learning, and knowledge graphs for robust IDP and outline future work on small language models and edge deployment.},
        keywords = {Intelligent Document Processing (IDP), Large Language Models (LLMs), Knowledge Graphs, BFSI, Hallucination Mitigation.},
        month = {June},
        }