proofessor reads your course material and points out what looks wrong: typos, code that fails, statements that don't hold, etc. It highlights these errors and suggests changes which in many cases can be implemented by simply clicking a button. However, the decision on what to change is left to the user.

What happens to your upload

Only the text is extracted from your files, so images, audio and video are not looked at. If the files contain Python or R notebooks, the code is executed in a sandbox with no network access to catch anything that fails. A local large language model then reads each document and reports what looks wrong and how it would fix it, and every finding and suggested fix is checked against your file before it is shown to you. Finally, each file gets a report page with its findings, where you review them and decide what to do with each one.

Where your material goes

The app runs on a VM controlled by the University of Copenhagen. This VM has a GPU which hosts Gemma 4 31B, a local open model that is used for reviews and suggestions. Since there are no external calls to LLM providers, you can upload anything you'd like. Your material is kept for 30 days, then deleted so that we don't need extensive data protection procedures.

Source for example courses

The example reports were made from courses written and taught by the Center for Health Data Science (HeaDS) at the University of Copenhagen, as published in August 2026.

The two example courses from HeaDS were evaluated by a state-of-the-art LLM (Fable 5.1) to list every major error: 14 in the Python course and 42 in the R course. We also listed passages that look wrong but aren't, to count false alarms. Each model was scored on how many of the known errors and false positives it found.

Results

Model	Python		R
	errors	false alarms	errors	false alarms
Gemma 4 31B open, local hosting	14/14	0/7	37/42	2/12
Fable 5.1 closed, external API	14/14	0/7	42/42	0/12
DeepSeek V4 Flash open, external API	14/14	0/7	41/42	3/12
GPT-5.6 Luna closed, external API	11/14	0/7	–	–
GPT-5 nano closed, external API	4/14	1/7	–	–

We also tested whether sending the same document more than one time helps uncover more errors. Each single Gemma 4 run found all 14 Python errors and 35–36 of the R errors, against 37 with three runs combined. DeepSeek showed the same pattern, so one run is usually enough.

While Gemma 4 doesn't reach a perfect score, it gets pretty close. Since it has the great advantage of being locally hosted, it is the default model used for this version of the app.

For each account we store a username, the password as a salted scrypt hash (never the password itself), and an encryption key that only your password can unlock. Your uploads, reports and accepted fixes are kept for 30 days and then deleted. They are readable on the server only while a review runs and until ten minutes after you last open them; the rest of the time they are encrypted with your key, so nobody else, including the operator, can read them once you have chosen your own password. The operator also keeps an activity log for 180 days: when you sign in, how many reviews you run and on which model, the types and sizes of your files, whether the reviews finished, and when you view or download a report. The log never contains file names or content.