Developer Needed for OCR Data Extraction using Machine Learning

Overview: The goal of our project is to create a software-based process for extracting text information from a variety of types of invoice documents (PDFs, scanned images (TIF, JPG, etc.)) that is based on an understanding of the document and the text on it instead of using fixed positions, etc. The end goal is to have this capability to be able to process any invoice that is presented to it. I see this effort being conducted in phases based on the success of each phase. The phases are as follows: 1. Upfront discovery call(s) with selected developer 2. Invoice Face Page Only proof of concept (see below) for US / English language invoices only 3. Invoice Face Page Only for other languages 4. Invoice Detail for US / English language invoices only 5. Invoice Detail for other languages Our business involves processing invoices so we require highly accurate invoice data. We have built a significant amount of invoice parsing technology covering a defined industry space but all dealing with actual data files, not OCR. This capability is needed to expand out to all industries as well as to automate internal processes involving managing invoice files. We believe we will implement this technology as part of a multi-step process within in our application to function as follows: 1. Invoices files will be transmitted to us in the following ways and will be placed into a file system a. SFTP b. AWS S3 c. Email attachments 2. This service will open each invoice file and perform its extraction process. The text information and other metadata (i.e. confidence levels, etc.) will be stored in a database table 3. Our application or this service will perform validation on the data extracted to determine its suitability to support our needs. Examples of validation will include things like the following: a. Match of extracted account number to our table b. Sum of various numeric fields (i.e. balance forward + new charges = total amount due) 4. For fields that cannot be validated through external means we will set some arbitrary threshold on the confidence level provided in the metadata and test actual data to determine how well this works 5. Any item failing validation will be set up in our application for a human user to review and fix. We will build this into our application UI Phase 1/2 Details: As part of the Phase 1/2 proof of concept we would like to start off with 1 or 2 discovery calls with the selected developer to better understand how the optimal technology (machine learning) for this task functions, the developer's experience / expertise in this area, etc. These will be paid discovery calls. Next, we want the selected developer to create software code to perform what is described below. - This proof of concept will be limited to a few invoices (5) to reduce the time to execute the project - Take scanned image files and do the following programmatically: ○ Extract defined fields based on learning what they are § Extract the following information (note not all fields may be provided): □ Vendor □ Account Number □ Invoice Number □ Invoice Date □ Bill Period From □ Bill Period Thru □ Due Date □ Customer Address □ Vendor Remit Address □ Balance Forward Amount □ Total New Charges □ Total Amount Due ○ For the sample images, I have defined where each of these field data are located § Again, we are looking for technology that understands the data and doesn't use position-based means of locating the data, but also realize that you may need information to train the model ○ Output data in a text file with following headers delimited by comma (i.e. .csv): § Vendor § Account Number § Invoice Number § Invoice Date § Bill Period From § Bill Period Thru § Due Date § Customer Address § Vendor Remit Address § Balance Forward Amount § Total New Charges § Total Amount Due