Margin

Writing is central to computing careers, but feedback on design documents, bug reports, and READMEs is slow, uneven, and expensive. In a class of 200, TA feedback can take weeks and tens of hours, and by the time it arrives students have moved on.

Margin is named for the margin of a draft, where a good reviewer’s comments have always gone. It is a retrieval-augmented generation pipeline that produces rubric-aligned, criterion-level comments rather than grades. It runs on small language models hosted on institutional hardware, so student work never leaves campus and there is no per-use API cost, building on the local small-model pipelines first developed for QuickTA.

In blinded, within-subjects studies, students rated the local model’s feedback 5.73 out of 7 in CS2 (n = 176), against 5.48 for GPT-4o and 3.62 for TA feedback. In Operating Systems (n = 80) the local model again led, and the pipeline processed the cohort in 1.8 hours instead of about 40 hours of TA grading and training.