Paragon Pipeline
An experimental OCR/NLP project for extracting data from Polish receipts. A pipeline concept and model research, without a production deployment.
- My role
- Pipeline design, model experiments, data preparation, classification and information extraction.
- Technologies
- Python / Jupyter / PyTorch ecosystem / NLP
Context
A receipt is a difficult document: irregular formatting, abbreviations, OCR noise and data that cannot be reliably parsed with simple regexes.
The project
I built a pipeline concept: ParagonOCR reads lines, ParagonNER classifies them and ParagonPIE extracts product information.
My contribution
Independent problem research, model experiments, pipeline stage design and documenting how the repositories connect.
Approach
The pipeline decomposes the problem into smaller stages: image → text lines → line classification → filtering → information extraction.
Challenges
- Splitting a complex ML problem into measurable stages.
- Handling Polish language, abbreviations and receipt formats.
- Thinking about the full system, not just a single notebook.
A similar challenge in your business?
Explore the possible scope and what to share when starting a conversation.