From Unstructured Documents toIntelligent Insights: An End-to-End AIPipeline for Global Tax Compliance
DOI:
https://doi.org/10.5281/zenodo.21358569Anahtar Kelimeler:
Document AI- Tax Compliance- Large Language Models- Natural Language Processing- Electronic InvoicingÖz
Over 100 countries now mandate electronic invoicing, with Turkey leading as an early adopter generating over two billion documents annually. For multinational enterprises, managing tax compliance across heterogeneous formats presents severe operational challenges. This study introduces an end-to-end AI pipeline that transforms unstructured invoices (PDFs, images, CSVs) into compliant data within a unified architecture.
The pipeline comprises three integrated components: First, the Smart Input Module leverages an ensemble of multimodal models (Gemini Flash, Nova Lite, GPT-4.1), achieving 98% accuracy on legally critical fields. Second, the Cloud Processing Layer utilizes a domain-specific XSLT deduplication technique to yield a 63% storage footprint reduction while preserving digital signatures, extending infrastructure runway to over five years. Supported by RabbitMQ and Kubernetes, this layer scales peak throughput from 32 to 200 transactions per second. Finally, the Sovi Intelligence Engine employs an agentic, retrieval-augmented generation (RAG) architecture to enable natural language querying across five languages, achieving a 94% valid SQL generation rate. Currently in production, this system demonstrates the viability of fully automated, AI-driven tax compliance.
Referanslar
Billentis, "E-invoicing/e-billing: Digitisation & automation," Market Rep., 2023. [Online]. Available: https://www.billentis.com/
Y. Xu, M. Li, L. Cui, S. Huang, F. Wei, and M. Zhou, "LayoutLM: Pre-training of text and layout for document image understanding," in Proc. 26th ACM SIGKDD Conf. Knowl. Discov. Data Min., 2020, pp. 1192–1200, doi: 10.1145/3394486.3403172.
Y. Huang, T. Lv, L. Cui, Y. Lu, and F. Wei, "LayoutLMv3: Pre-training for document AI with unified text and image masking," in Proc. 30th ACM Int. Conf. Multimedia, 2022, pp. 4083–4091, doi: 10.1145/3503161.3548112.
D. Wang, N. Ravi, C. Arber, S. Bhatia, and A. Chakravarthi, "DocLLM: A layout-aware generative language model for multimodal document understanding," in Proc. ACL, 2024, pp. 2551–2568.
L. Li, Y. Zhang, and L. Chen, "A survey on large language models for recommendation," World Wide Web, vol. 27, no. 5, pp. 1–35, 2024, doi: 10.1007/s11280-024-01291-2.
A. Plaat, T. Muller, and J. van den Herik, "Agentic large language models: A survey," J. Artif. Intell. Res., vol. 82, pp. 345–412, 2025, doi: 10.1613/jair.1.16853.
M. Pourreza and D. Rafiei, "DIN-SQL: Decomposed in-context learning of text-to-SQL with self-correction," in Adv. Neural Inf. Process. Syst., vol. 36, 2023.
A. Plaat, A. Wong, and J. van den Herik, "Multi-step reasoning with large language models: A survey," ACM Comput. Surv., vol. 57, no. 3, pp. 1–38, 2024, doi: 10.1145/3717843.
Langfuse GmbH, "Langfuse: Open source LLM engineering platform," Software documentation, 2023. [Online]. Available: https://langfuse.com
J. Li et al., "Can LLM already serve as a database interface? A big bench for large-scale database grounded text-to-SQL," in Adv. Neural Inf. Process. Syst., vol. 37, 2024.
İndir
Yayınlanmış
Sayı
Bölüm
Kategoriler
Lisans
Telif Hakkı (c) 2026 Muhammed Şara, Ahmet Ak, Ayhan Boyacıoğlu, Yunus Taştutan (Author)
Bu çalışma Creative Commons Attribution-NonCommercial 4.0 International License ile lisanslanmıştır.
Makale gönderen yazarlar, Aintelia Science Notes Journal'da (ASNJ) yayınlanmak üzere kabul edildikten sonra çalışmalarının telif haklarını saklı tutarlar. Göndererek yazarlar şunları kabul ederler:
- Telif Hakkının Saklı Tutulması: Yayınlanan makalenin telif hakkı yazar(lar)da kalır. BAUDER Press'e, çalışmayı yayınlama, çoğaltma, dağıtma ve arşivleme konusunda münhasır olmayan bir lisans verilir.
- Lisanslama: Makale, CC BY-NC 4.0 lisansı altında yayınlanır ve üçüncü tarafların, yazar(lar)a ve dergiye uygun şekilde atıfta bulunulması koşuluyla, çalışmayı ticari olmayan amaçlarla paylaşmasına ve uyarlamasına izin verir.
- Yazar Hakları: Yazarlar, derginin orijinal yayın yeri olarak belirtilmesi koşuluyla, makalelerini herhangi bir ortamda (depolar, tezler, konferanslar, öğretim materyalleri) kısıtlama olmaksızın depolama, dağıtma ve yeniden kullanma özgürlüğüne sahiptir.