OneLook.

Diabetic retinopathy (DR) is graded on a 5-stage severity scale that normally requires an ophthalmologist reading a fundus photo. OneLook (built as RetinAI) grades both eyes from uploaded photos directly, no specialist required for the first pass.

It's a single-purpose classifier, not a general diagnostic platform — DR grading is the only thing the model does.

A DenseNet121 backbone (ImageNet-pretrained, top removed) feeds into global average pooling, dropout, and a 5-way softmax head — trained to output one of No DR, Mild, Moderate, Severe, or Proliferative DR. Left and right eyes are classified independently by the same model.

The app's own UI copy states the model runs at roughly 85% accuracy — the only accuracy figure that exists for this build; there's no separate benchmark or held-out test report backing it further.

Each eye's grade fills a DOCX report template alongside patient fields and the source images, plus a stage-specific canned recommendation — for example, Proliferative DR triggers a note that immediate treatment is essential to prevent vision loss.

The DOCX is converted to PDF (docx2pdf), rendered to an image for in-app preview (pypdfium2), and offered as a download — which is also why the report pipeline is Windows-bound (docx2pdf's conversion path depends on pywin32).

Built and shipped in a single session in August 2023 as a competition entry. Model weights are hosted externally rather than committed to the repo, so running it locally means downloading them separately before the first launch.