Assessment AI

While employed at Cosairus, I developed an API service using Django and spaCy which could be sent text, which it would analyze and search in for sensitive medical information, such as mentions of substance abuse and STDs. These findings would be reported back to the caller.

The service was then rewritten as a headless CLI application that processes a set of files. After this was completed the project was handed back to me and I implemented several additional features. These included accessing files from an API, extracting text automatically from PDF files, storing results to a SQL database, creating and uploading an annotated version of the PDF file with the sensitive information highlighted for review, detecting commingled records from multiple patients, and detecting release authorization forms.

I also created a system where human QA reviewers can provide feedback on the AI’s findings. This feedback was saved and displayed in an administration application, where it can be approved or rejected by a manager. After a process has received approved feedback, it can be submitted for retraining.

Once this happens, a Python service retrieves the information from the database about the AI’s initial findings and the approved feedback, redownloads the searchable PDF file to extract its text, and generates newly refined training data in an automated process.