Invoices with Azure AI

📌 Project specifications

Goal:

Explore the potential of Azure AI in automating invoice processing within BIM4Production, and investigate how this technology could support future AI integration across the company’s ecosystem.

Scope:

• Design and develop a prototype to extract structured data from invoices using Azure AI.
• Benchmark its performance against the current external solution, which has a ~30% error rate.
• Build organizational knowledge around Azure AI and identify requirements for scalable integration.

Timeline:

15 Weeks

🖥️ Technical Specifications

Technologies, tools and frameworks used:

• Azure Document Intelligence service
• C#

Model architecture:

The solution progressed from Azure’s prebuilt invoice model to a custom-trained model tailored to company-specific invoice layouts. A feedback loop was introduced for retraining and improved accuracy. The architecture was supported by C4-model diagrams to ensure future scalability and maintainability.

🎯 Results

The developed prototype, based on Azure Document Intelligence, managed to extract invoice data with a notably lower error rate than the company’s current external tool. By training a custom model and embedding domain-specific validation rules, we were able to reduce the number of incorrect predictions — improving both reliability and maintainability.

Alongside the prototype, I delivered an architectural blueprint for a scalable Azure-based solution. This design incorporates feedback loops, configurable logic, and integration points tailored to BIM4Production’s ecosystem. Together, the working prototype and architecture provide a solid foundation for future AI-driven improvements to their business processes.

💭 Reflection

This project offered me a rare opportunity to explore both applied AI and the broader system thinking that underpins its success in practice. While the prototype was practical in nature, the process of building, training, and validating models in a real business context deepened my understanding of core AI principles — like model generalization, data quality, and the challenges of domain adaptation.

More than just coding, this internship helped me think critically about how AI systems can be responsibly integrated into real-world workflows. It also strengthened my interest in the theoretical underpinnings of machine learning — something I look forward to exploring in much more depth during my upcoming Master’s in AI. For me, this project confirmed how much I enjoy working at the intersection of research, development, and meaningful application.