Back to Projects
Professional ProjectAI/ML

CNN-based Document Field Extraction

Timeline: 2019
Role: AI Engineer & Solution Architect

Overview

First AI solution integrated at Springtime Technologies - CNN-based field extraction system with custom embedding layer for automated invoice processing.

Challenge

No existing AI capabilities in company. Needed to extract multiple fields from invoices (amount, number, date, etc.). Required accuracy better than manual processing. Customers unfamiliar with AI technology, demanded traceability and explainability.

Solution & Approach

Designed and integrated ResNet architecture in TensorFlow with custom embedding layer. Embedded OCR and invoice metadata to simplify learning task and improve accuracy. Built end-to-end training, testing, and deployment platform from scratch. Developed traceability and explainability features to address customer concerns. Integrated AI solution into worldwide Invoicetrack platform.

Outcome & Impact

Achieved 10% lower error rate than manual clerk processing with similar automation rate. Deployed to production environment. Integrated into global Invoicetrack solution. Adoption delayed due to customer AI unfamiliarity in 2019. Developed traceability and explainability features to build customer trust. Established company's first AI capability and integration practices.

Technologies Used

TensorFlowKerasResNetPythonOCRCustom embedding layers

Key Highlights

  • Company's first AI capability
  • Achieved lower error rate than manual processing
  • Custom embedding layer for OCR and metadata
  • Developed traceability features for customer trust
  • Deployed to production
View All Projects