Skip to main content

Model flow

The flow of an H2O Document AI model from creation to deployment and consumption can be summarized in the following sequential steps (discussed in the below sections):

H2O Document AI architecture model. The process is broken down in the following steps.

Step 1: Ingest

Upload your documents to H2O Document AI using the Document AI web interface or API. H2O Document AI lets you handle a wide variety of documents, including:

  • Image scans (faxes in PDF or other formats, pictures with text, and non-editable forms)
  • Documents with embedded text which have text and layout metadata (PDF docs, Word docs, HTML pages)
  • Documents with regular text “left to right/top to bottom” (CSVs, emails, editable forms)

Step 2: Pre-process

Pre-process documents before training with a set of state-of-the-art computer vision and NLP product features. Pre-processing includes support for:

  • Recognizing and handling embedded text
  • Recognizing and handling logos
  • Page orientation resolution
  • Deskewing
  • Cropping
  • Text formatting optimization
  • Color binarization
  • Addressing input PDF quality challenges

Step 3: Labeling

Add, improve, and validate document labels:

  • Integrates with common label formats
  • Provides advanced options for validating labels against scored documents and determining labeling sufficiency

Step 4: Train models

Select the training data set within H2O Document AI, and it will automatically learn the document and create models.

  • Language understanding and layout recognition using learning based on deep learning, transformer architectures, and machine learning
  • AI-ML engine that uses multiple computer vision and NLP algorithms for diverse AI tasks
    • Entity recognition
    • Document and page classification
    • Form understanding
    • Grouping & set identification

Step 5: Post-process

Post-process to ensure consistency, accuracy, and organization of scored documents. H2O Document AI lets you perform a range of customized post-processing jobs that use AI algorithms vs. rules to ensure high-quality predictions and insights.

  • Organizing prediction sets
  • Confidence and probability measures
  • Datatype standardization – date, times, currency codes, international numerical formats, locations

Step 6: Deploy models

Publish models into your cloud or on-premises environment of choice. Integrate models into existing systems, processes, and applications via APIs or JSON documents.

Step 7: Consume

After deploying your models, you can:

  • Consume the model through business apps
  • Store data
  • Batch score in real-time to obtain predictions

Feedback