Intelligent document processing with AI

ML-based PDF infomartion extraction system with storage and search functions.

Currently pretrained for Hungarian EKR documents (some official national contracts), but you can train 6 different models with your own data.

Services:

Backend

• upload pdfs
  • store pdf file in AWS S3
  • exract text with Tesseract OCR
  • exract information using the model service
  • store text data in Elastic Search
• search
  • return pdf text data by query (match/levenhstein/regex/...)
• download
  • download pdf file by filename

Tech: JavaScript, Express.js, Pdf-Poppler, Tesseract-OCR, Elastic Search, AWS S3

Model

The backend can run any .py and .ipynb files as with the excepted input/output formats

• predict

  • batch text information extraction with CRFSuite ML model (Conditional Random Fields)
  • many other models have been tried, but those reached lower accuracy for this amount of data

• train

  • todo

• tested models (dataset):

  • Custom neural networks:
    • Embedding + bi-LSTM
    • Embedding + bi-LSTM + LSTM
    • Embedding + bi-LSTM + LSTM + CRF
  • Bert
  • XGBoost
  • CRFSuite

Tech: Python, Flask, Keras, PyTorch, Bert, XGboost, PyCRFSuite

Frontend

• draft

Tech: JavaScript, React