WK·004
Gradience
An assisted grading workflow that was probably right about the direction, but early for the market and the models.
41 commits · 15 active weeks
latest 33c2637 · 21 Apr 2025
Overview
Gradience, originally Assisted Grading, explored how AI could support the assessment workflow without removing the lecturer from the process. The idea was to help with assignment setup, rubric generation, document upload, criteria-based evaluation, editable feedback, and results across a cohort.
The product assumption was simple: grading is slow and inconsistent, but the answer is not to hand judgement to a model. The answer is to make the repetitive parts clearer and faster so the human examiner has more time for the judgement that actually matters.
The walkthrough
Actual footageThe real Gradience app, start to finish in about half a minute. Drag the timeline to move through the six stages, or let it run. Recorded on a demo account with sample data, so nothing here is a real student.
WK·004 · one continuous cut of the real Gradience app, speed-ramped to each stage's essential moment · demo account, sample data, no real student work
Context
We designed the workflow, produced the UI direction, built around rubrics, and pitched the idea. On paper, it made sense. In practice, the timing was early.
Irish higher education was cautious about AI in assessment, and rightly so. Academic integrity, explainability, auditability, and trust are not footnotes in this domain. At the same time, the models were not yet good enough at reliably reading longer submissions, understanding the assignment ask, and mapping evidence to a rubric in a way a lecturer could stand over.
What changed
The useful part of the design was the rubric-first shape. A score without a visible criterion is just a number with confidence attached. Gradience tied evaluation back to named criteria and editable feedback, which is the only way a lecturer can inspect the system rather than just receive its opinion.
The bad bargain in automated grading is speed at the cost of trust. Nobody needs faster marking if every result creates an argument. The better target is a workflow that makes standards explicit, keeps the human in charge, and leaves an audit trail for why feedback was given.
The build
The system shape was deliberately staged: describe the assessment, draft or select a rubric, upload submissions, evaluate against criteria, review feedback, and aggregate results. The model work was only one part of that. The institutional plumbing matters just as much: roles, courses, permissions, reused rubrics, and records a lecturer can defend later.
Looking back, the idea feels more plausible now than it did then. The models are better. The governance conversation is more mature. The market still needs care, but the technical ceiling is very different.
Notes from the archive
I do not think Gradience was a bad idea. I think it was early. The market was cautious, the models were not ready enough, and the trust problem was bigger than a demo could solve.
If I built it again, I would start even more deliberately with governance and auditability. The model would be a feature inside the assessment workflow, not the headline. That is less exciting on a pitch slide and much closer to how institutions actually buy systems.
Evidence
| repo | Private repositories (grading API and UI), reviewed for this archive | repo-reviewed |
| commit | Latest reviewed API commit · 33c2637 | 2025-04-21 |
| video | Walkthrough recorded from the live app (demo account, sample data) | source repo |
| note | Multi-tenant by design: institutions, courses, and roles (admin, university admin, grader) | architecture |
Technical detail
Other exhibits
| WK·001 | Alph | An open source CLI for MCP setup, built because editing the same JSON in five different places is a poor use of anyone's evening. | Terminal |
| WK·002 | AskHuman | A hosted MCP loop for agents that need a human decision instead of another confident guess. | Decision loop |
| WK·003 | Rhya | A live wellness site for a real business, built close enough to the stakeholder that vague feedback was never going to survive long. | Capture |
| WK·005 | WatstheStory | A personalised audio briefing system for WhatsApp, built because a short spoken update sounded better than opening ten feeds before breakfast. | Live |