Use Open Source to Extract Data

[jillcfoley1]

As a system, I leverage an open-source OCR tool to automate data extraction so that the product team can evaluate the tool's effectiveness compared to AWS Textract.

AC:

• Explore and evaluate open-source OCR tool options, focusing on essential capabilities such as character recognition accuracy and ease of integration. Identify the highest-merit open-source tool for initial experimentation, prioritizing efficiency, measurable outcomes, and automation potential.
• There is a second application that uses the open-source tool, in addition to the AWS Textract application.
• The new open-source tool is integrated into the new application.
• The team has designed and executed side-by-side comparison testing across both applications and has an understanding of the accuracy of both. Evaluate measurable outcomes such as accuracy rates, processing speed, and error handling capabilities.
• The team has a documented understanding of the comparative effectiveness, measurable savings potential, and any efficiencies gained through the use of an open-source tool.

Technical Details
Technical Stack Considerations:

• Identify compatible OCR libraries (e.g., Tesseract, PaddleOCR, or Textract alternatives).
• Ensure selected tool supports required file types (PDF, TIFF, PNG, JPG).
• Implement logging and structured reporting for extracted data, errors, and confidence scores.
• Where possible, containerize the OCR service for consistent deployment across environments.

Data Pipeline Integration:

• Ingest source files via standardized input folder structure, API call, or message queue.
• Normalize outputs (structured JSON, XML, or CSV) to align with product team’s existing comparison scripts.
• Implement error handling for failed pages, unsupported formats, and language mismatches.

Testing and Validation:

• Create a standardized test corpus representative of real-world data.
• Compare outputs using measurable criteria:
• Character-level accuracy
• Field-level completeness
• Confidence score variance
• Processing time per page
• Use automated comparison scripts to generate side-by-side accuracy reporting.

Reporting Artifact:

• Deliver structured evaluation report covering:
• Tool selection rationale
• Accuracy comparison (open-source vs AWS Textract)
• Processing performance (speed and failure rates)

Hypothesis: OCR validation: We believe that integrating a merit-based open-source OCR tool can streamline operations and reduce waste by replacing AWS Textract—delivering equivalent or superior accuracy without compromising essential performance standards.