Automating Document Verification Demo

I want to quickly provide some important basic concepts that will help you all understand what we are trying to achieve with our platform. At a high level, Hyperscience is trying to turn human-readable data into actionable machine-readable data downstream. Hyperscience will ingest these documents, attempt to classify them, identify the pertinent data points from these documents, and then extract those data points.

This is all based on a unique aspect of Hyperscience. Out of the box, you will be able to set your target accuracy output SLA, and that is essentially your north star. Everything is built around this. If the machine at any point in these steps is unsure that it won’t meet that SLA, it is going to raise its hand and ask for the human to assist. This is what we call the human-in-the-loop experience.

With Hyperscience, you can also add business rules to this experience to help enrich the data. For example, if you want to set up an API call to validate an address or tap into a database to validate an account number. Additionally, we think of documents in a few different buckets: structured (template-based documents where we know where the data is), semi-structured (documents where we know the fields we want to capture but not their exact location, like pay stubs or invoices), and unstructured (freeform documents like doctor’s notes or emails).

We are going to jump right into the demo. I am going to take you through a high-level overview of the Hyperscience Platform where we are trying to solve a very manual process for our customer today. Currently, their customers send in documentation to apply for a loan. This requires people to go into these documents, look them over, grab the pertinent information, and then manually enter it into their application system downstream.

In this story, we have Thomas Edison applying for a loan. He is emailing in some documents. First, we have an application form with a lot of handwriting, which Hyperscience handles better than anybody in the world with our proprietary engine. This is a structured document where you can draw boxes around fields. We then have a W-9, another structured document submitted upside down with a mixture of handwriting and machine text. Hyperscience processes both with the same engine. The last document is a pay stub, giving you an idea of what Hyperscience can do with semi-structured documents. We know the data points we want to capture, but we cannot template this out because of the variation.

Let’s jump into the platform. This is the Hyperscience platform, your single pane of glass for everything from building models and layouts to administering the platform and reporting. It is a SaaS-based application, but there is an on-premise version and the ability to deploy in a private cloud. We have a specific flow set up for those loan documents. We have the email listener, so anytime the document gets sent into that email, Hyperscience will ingest it.

The machine will try to classify the document, identify the pertinent data points, and then transcribe those points. If the machine is not confident it meets that target SLA accuracy, it will bring a human in the loop. In this particular flow, I want Hyperscience to look for specific data points across any of the documents, create a case to group them together, and allow me to validate that I have all the data points needed to process the loan application. I am looking for account numbers, pay date, gross earnings, social security number, and address.

These documents have already been classified without needing my help: the loan application, the W-9, and the pay stub. How does Hyperscience classify these? You have a library of all the different types of layouts and document types. For structured documents, I am simply templating out the document, drawing fields around the boxes I want to capture. When setting up these layouts, you are also giving in a data type. Hyperscience has an out-of-the-box natural language processing model trained on billions of data points. You use data types to give context to the machine for what is expected in each field.

For pay stubs, which are semi-structured, we simply tell the machine the types of fields we want (address, name, etc.) and submit documents to create a training data set to build a predictive model. Now, let’s play the role of the data keyer. The machine has already classified the documents but has raised its hand looking for help with the pay stub manual identification step. With our predictive model, Hyperscience has captured a lot of these fields in an automated fashion, but there are a couple of fields the machine is raising its hand on. I can see the period end date on the document. All I need to do is highlight and click over the data point to capture that field. This not only completes the submission but provides data points to improve the model in the future.

We have done the machine classification and identification steps. Now we are at the transcription step. The machine is raising its hand again. With transcription, the machine brings you right to the particular field you need to help with. I see Thomas Edison’s name, and I have the context that it is an employee name. I simply type what is there. That is the only piece I needed to help out with. The user experience is meant to be quick, easy, and efficient. Taking a fully manual process and automating most of the work combined with an efficient user experience typically results in faster throughput times.

We have one more step: confirming that we have the right data points to approve or deny the customer. Hyperscience brings me to those data points. We are pulling the social security number. I can click on that, and it brings me to the field. Even with the way this was written—where the six is on the line and the eight and one are outside the box with a mistake in between—Hyperscience is able to grab that data and validate it. We have the pay stub with the pay date and gross earnings. For the W-9, it doesn’t look like Hyperscience found a value for the account number. I can see why; there is nothing in that field. I am going to reject this application and move on. If I reject that application, I then create an automated process to contact the customer about the status.

Hyperscience did most of the work. It was 92% done by Hyperscience. We only had to help with one field across three documents for transcription. Everything is sent downstream through a JSON. You can see how well Hyperscience handles handwriting and machine text. For example, it knows we are looking for a name and ignores mistakes. It knows the format of an address and looks at the entire context of the field, not just character by character. The social security number data type helps out because it looks for a specific amount of digits and ignores background noise. Finally, Hyperscience can normalize data, such as converting a written date into a specific number format for downstream systems without needing human help.