The client is a Leading Online Education Provider Company that is affiliated with more than 300 schools in the United States and receives more than two million applications from potential students annually. The application requirements for each individual school associated with the online services provider is different and each application is submitted with several supporting documents, such as: Birth Certificate, Prior Schooling Transcripts, Proof of Address, and so on. During peak enrollment times, like summer and early fall, the client was employing, training, and maintaining 120 part-time enrollment officers on top of their 35 full-time employees to spend their full days, plus overtime, processing enrollment applications. Specifically, the enrollment officers had to manually process all the documents, adjust all images submitted to meet an exact set of image guidelines and then label each application as accepted or rejected based on the fact that all documents and images submitted were the appropriate formats and met the list of image requirements. Each of the clients associated schools had fixed guidelines for document approvals and are listed as follows:
- Documents should be properly aligned in portrait mode
- Documents cannot be skewed in any way
- Document Images should have no visible background 4. Documents should be of standard US letter size (8.5×11 in.) 5. Documents submitted must be listed as a specified acceptable document for each category on the application
With 85% of the document images submitted in the applications being images taken via mobile device camera it was almost guaranteed that every image taken by this type of camera was going to be skewed, rotated the wrong way, and/or with a visible background present. All of these factors conflict with the application document image guidelines, causing the enrollment department to spend countless hours adjusting the images, or rejecting the application, causing their customers frustration and even more employee time as they would have to repeat the process all over again when the customer submitted a revised application.
Post Process Analysis, we concluded that we would need to combine machine learning, computer vision, and RPA together, to create a precise, consolidated automation for the client. The need for this customization stemmed from the realization that each document submitted was so unique that general utility applications and nothing currently available in the market was capable of carrying out the automation without further support. To tackle the design of this cutting-edge, customized automation, we split the problem into four categories.
- Document Image Automatic Rotation and Skew Detection:Deep-Learning-Based Machine Learning Models were trained with historic images that were manually processed by the enrollment department in previous years. Computer Vision Algorithms were used to rotate images and remove their identified skew to meet the document image guidelines.
- Document Image Automatic Cropping:Machine Learning and Computer Vision-based edge and corner detection algorithms were applied to locate the corners and edges of the document in the image accurately. The image is then cropped precisely up to the document edge to remove any part of the image that is not part of the document photographed.
- Document Classification:Machine Learning-based document classifiers were created and trained using the client’s large database of correctly labeled and accepted or rejected documents and application submissions. The bot labels each document correctly, matches it to the list of acceptable documents for each category (ex. “School X” might only accept a utility bill or a bank statement as an adequate submission to provide proof of address. So, if a photo of a parent’s driver’s license is submitted, even though it has their resident addresses listed on it, it is not an accepted form by that school and is rejected) and then accepts or rejects the application depending on whether or not all the documents match the accepted forms for each application category.
- Document Intake and Upload Pipelines:Bots intake all the applications uploaded to salesforce using API’s and execute all the steps above as each is an independent python module in the RPA workflow. The processed documents and attached appropriate decision to either accept or reject is then uploaded to the school’s respective application.