Screenshot Topic Classifier

A three-class screenshot classifier (code, figure, text): a linear probe on frozen CLIP image embeddings, calibrated confidence routing with blind human review, and leave-one-origin-out evaluation.

What it is

A pipeline that sorts screenshots into three layout classes (code, figure, text) with a linear probe on frozen CLIP image embeddings. Calibrated confidence decides what is filed automatically; the rest goes to blind human review, and the live stream is monitored for drift.

The problem

How it works

Results / what I learned