AI-Assisted Marking: What It Actually Means for Academics
A practical look at how AI fits into the exam marking process — what it does well, where it falls short, and why the examiner stays in control.
ai-markinghigher-educationworkflow
Blog
Thoughts on AI-assisted marking, privacy in education technology, and product updates.
Privacy-First Grading: Why It Matters More Than You Think
Student exam data is some of the most sensitive data a university handles. Here's why privacy architecture should be the first question you ask of any AI marking tool.
privacygdprhigher-educationsecurity
Designing Rubrics That Work Well With AI Marking
The quality of AI-assisted marking depends heavily on the quality of the rubric. Here's how to write marking schemes that get the most out of automated assessment.
rubric-designai-markingworkflowhigher-education
Six thousand tests and counting: testing an AI-powered application
AEMS has over 6,600 automated tests. This post explains why that number is so high, what the tests actually cover, and what we learned about testing software that depends on non-deterministic AI models.
testingengineeringqualitytechnical
Building a memory system for examiner corrections
When an examiner corrects an AI-proposed mark, that correction contains information. AEMS records these corrections in a structured memory system that improves subsequent grading without retraining the underlying model.
memory-systemcalibrationtechnicalai-marking
Local-first architecture for student data
AEMS was designed so that student exam data never needs to leave the examiner's machine. This post explains the architectural decisions behind a local-first approach and why it matters for university adoption.
privacyarchitecturegdprtechnicalsecurity
Integrating with Canvas LMS: lessons from building a five-step grading wizard
Canvas is the LMS used by most Swedish universities. Building a grading workflow that operates within Canvas rather than beside it required navigating API limitations, permission models, and the reality of how examiners actually work.
canvas-lmsintegrationworkflowtechnical
Reading handwritten exams with vision models: what works and what does not
Vision-capable language models can read student handwriting, but the path from raw PDF to reliable text extraction is more complex than a single API call. A practical account of the challenges.
vision-modelshandwritingpdf-processingtechnical
Why a mechanics professor built an exam grading tool
The origin of AEMS was not a startup pitch or a grant proposal. It was a stack of 280 handwritten exams, a week of marking, and the realization that the process had not changed in decades.
origin-storyhigher-educationsolid-mechanicsworkflow