<aside> 📄
</aside>
Goal: Plan a solution to the problem and find a dataset with text data from OCR
Solution: This Kaggle dataset
TL;DR: Found a Kaggle dataset that contain images and OCR results of receipts. Planned and set up GitHub.
Full Notes:
Inspired by an AI engineer system design interview: “From a CSV of receipt image paths, how can I build an ML system to extract invoice numbers?”
My potential solution:
Clean images (align, orient, etc.) → OCR → Extractions by heuristics → Extractions by LLM → Human-in-the-loop verification / extraction
Selecting a dataset:
I wanted to focus on the LLM engineering part, so I found a dataset that contains OCR results for simplicity. Check out my 📝 documentations and notes on the dataset here. It contains images, OCR’ed results, but is lacking ground truth labels for invoice numbers.
If my goal is to train and finetune a model to extract invoice numbers, I will need ground truth labels. My immediate next steps are to create a data labeling tool to streamline data labeling, and begin labeling data. Labeling data will help me design heuristics to extract data.