The Company

Infrrd is a pure-play intelligent document processing (IDP) software company. Launched in 2016, Infrrd began with a mission to use new advancements in AI around computer vision, natural language processing (NLP), and predictive algorithms to create new business solutions. The company has focused on automating human tasks that require low-level cognitive capabilities, such as reading a document, entering data from forms, understanding contract clauses, and extracting data from images. Today, Infrrd offers an AI-powered data extraction platform for semi-structured and unstructured documents from document-intensive industries such as insurance and financial services and has grown to over 300 employees.
Infrrd positions its solutions in contrast to legacy capture companies that use optical character recognition (OCR) and templates, claiming to deliver “next-generation” OCR that does not require templates and can extract data from complex documents with high accuracy, where OCR alone has traditionally struggled. Recently, the company made waves by announcing a 100% data extraction guarantee (read our note about that).

The Technology

Infrrd IDP is a cloud-native, AI-driven platform designed to extract data from any document that contains text or tables and provide formatted data to other business applications that depend on structured data; for example, accounts payable, claims processing, new customer onboarding, loan origination, etc. The company can provide private cloud deployments.

The first step is the creation of an IDP data extraction model. Infrrd offers dozens of pretrained models for common document types such as invoices, receipts, insurance forms, tax forms, statements, and more (see Figure 1). You start by selecting a pretrained model, then use your own documents to fine-tune the algorithm.
Additionally, new models can be created for documents the system has never seen before. This is accomplished through the Infrrd knowledge graph, a repository containing more than 500 fields currently understood by the AI engine. Infrrd expects this knowledge graph to expand rapidly.

Infrrd created a straightforward model-training process that does not require any knowledge of data science or AI. Define the data fields for extraction, upload a sample document, and train the model using confidence scores. Rinse and repeat until you reach the desired level of accuracy. Next, run the model in a production setting with the actual documents, and correct any errors using the Infrrd UI. The model is continually learning and improving.
Under the covers, Infrrd has built a strong AI engine using five key technologies:

1. NLP
2. Machine learning (ML)
3. Neural networks
4. Computer vision
5. OCR

The company developed its own AI in-house and was awarded patents on several key document AI methodologies, including confidence score algorithms, reading text from images, separating unique documents within a composite file, linking entity relationships in a document, and extracting data from invariant templates. In our briefing, company representatives promised more patents to come.

In particular, there are three IDP features where Infrrd excels. The first is table extraction, an essential feature for processing statements and invoices. Infrrd performs well on complex tables including those that are nested or have no borders (see Figure 2). The Infrrd model can also extract similar columns of data from different document types or variations. Underpinning this is the use of a dynamic graph convolutional neural network.
The second is the ability to extract data from unconventional images such as logos, banners, seals, and stamps. This feature is essential for use cases such as freight forwarding or mortgage processing, where documents contain such images. Infrrd developed a suite of deep learning models to handle the trickiest problems, such as poor image quality, incorrect orientation, or noisy backgrounds.

The third feature – and one that Infrrd says resonates the most with their customers – is Infrrd’s ability to handle millions of variable documents without templates. This is a direct benefit of deep learning pre-trained models and the knowledge graph.