A marking assistant that stays inside the campus network
A department needed marking support but could not send student work to a third party. We tuned an open-weight model and deployed it on their own hardware.
The constraint came before the requirement
A business school department marked 1,400 scripts a term across six modules. Marking support was welcome; sending student work to a hosted model was not permitted under their data policy, and no exception was going to be granted.
Open weights, tuned on their own rubrics
We tuned an open-weight model on four years of marked scripts and the department rubrics, deployed on two GPUs in the campus data centre. It drafts rubric-aligned feedback and a provisional band; the marker confirms or changes both, and the change becomes training signal for the next cycle.
Measured against double marking
Before rollout we ran the model against 300 scripts already double-marked by staff. Band agreement with the human markers came out at 96%, within the range the two human markers agreed with each other.
Markers writing less, judging more
Marking time fell by around eleven hours per marker per week, mostly in drafting feedback rather than deciding grades. No student work has left the campus network, which was the condition the project had to meet before anything else counted.
We were told this was impossible under our data policy. It turned out to be a deployment question, not a policy one.
ABHead of Department, Ashbourne Business School