PrelimMD

Front-desk staff should not retype a form the patient already filled, and they should not trust a read the software is unsure about. I started PrelimMD for that gap.

The problem

A clinic intake form comes back as a PDF, a scan, or a photo. Someone at the desk types it in again, and the second copy is where the digit changes. The person it frustrates is the one retyping, and the patient whose answer was legible on paper and wrong in the system. It was worth building a reader that also says when it is unsure.

The useful output is three things: the answers, a mark on the ones the reader does not trust, and a blank copy filled in. A fill with no mark is how a bad read gets signed.

How I built it

You upload a PDF, a scan, or a photo. The app finds the fields, reads the answers, and flags the ones it does not trust. It can also refill a blank copy of that form, so the output is a form again and not only a JSON dump of what the reader thought it saw.

The stack follows the three inputs. pdfplumber reads a PDF that already has text. Tesseract and OpenCV are there for the scan and the photo, where the page is an image and the fields have to be found before they can be read. FastAPI is the API. Next.js is the page the upload starts from.

On the zero-shot check in the limitations note, recall was 67.4 and precision 78.2 across 112 forms. Handwriting, on 49 fields, landed between 43 and 53 percent. Those figures are why the flag exists. A reader in that range should not get the last word.

Challenges and what I learned

The flag matters more than the fill. Handwriting on those 49 fields stayed between 43 and 53 percent, so a confident autofill would have been wrong about half the time on the hard fields. This stays a research prototype, not a clinical system. The blank copy is useful only next to the marks the reader does not trust.