AI-powered
data extraction
for invoices receipts ID documents any document

Cradl is created for development teams with advanced data capturing needs. Our users are typically developers in enterprises with high volumes of documents who want to take full advantage of machine learning technology to automate document processes.

A number of factors affect the accuracy of a trained model. In practice, the most important factors are the size and quality of the training dataset.

For documents like invoices and receipts, a well trained model can have field-wise accuracies over 99%. Furthermore, every prediction comes with a corresponding confidence which measures the probability that a prediction is correct. By filtering out low-confidence predictions based on a threshold, you can increase the accuracy by making a trade-off between accuracy and automation.

For example, assume you have a model trained on invoices, and that you want 99.8% accuracy for total amount predictions. Through experimentation, you may find that if you set your confidence threshold to 0.9, then 75% of all total amount predictions have a confidence level above 0.9, and you observe a 99.8% accuracy for these predictions. The remaining 25% can be sent through a manual verification step.

The higher the variation in the documents, the more training data is required. Training a model to process receipts from all over the world requires significantly more training data than if you are training a model to read national ID cards.

We recommend training on at least 5 000 documents. The training data should be as similar as possible to the data you want the model to process. Training data can often be automatically extracted from a database, archive, etc. so that manual document labelling is not required.

No. As long as you can use our REST APIs, Cradl AI takes care of the heavy lifting.