EdTech
AI-first student management portal
A portal where custom models mark the work and draft the feedback, and a teacher signs it off.
- Less time spent marking
- 68%Less time spent marking
- Of proposed marks accepted unchanged
- 94%Of proposed marks accepted unchanged
- Submissions graded in the first term
- 9,400Submissions graded in the first term
- Answer formats graded on separate paths
- 3Answer formats graded on separate paths
The problem
Teachers were losing their evenings to marking, and feedback reached students long after they had moved on. An earlier attempt at automated grading had been switched off after a fortnight because it marked written answers with the same logic it used for arithmetic.
The approach
A portal built around custom grading models, generated feedback and cohort analytics, with a teacher confirming every mark before it counts.
Grading is not one task
Grading looks like one job and is at least three. A numerical answer is a comparison against a key, with credit for method. A short written response needs semantic judgement against a rubric. A structured piece of work, say a proof, some code or a lab write-up, carries partial credit spread across several steps. Putting all three through a single model is the fastest way to lose a teacher in the first week, because it will be confidently wrong in a way that is obvious to them and invisible to you.
So the portal routes by question type before it grades anything. Each route has its own model, its own rubric format and its own confidence behaviour. Numerical marking is close to solved. Structured work is the one we still watch.
The teacher stays in the loop
The teacher is in the loop by default rather than as a review step bolted on afterwards. The model proposes a mark together with the reasoning behind it, the teacher accepts or overrides, and nothing reaches a student until a person has agreed to it.
The overrides are the useful signal. Each one is stored with the original proposal and the rubric item it was judged against, and they feed back into tuning. Across the first term about 94% of proposed marks went through unchanged, but the 6% that did not were worth more than the rest put together, because they clustered, and the clusters pointed at specific rubric items the model was reading too literally.
The analytics came out of the structure
Analytics were close to free. Once marking is structured data rather than a spreadsheet of totals, a question like "which rubric items is this cohort weakest on" is a query rather than a project.
That changed what teachers asked for. The original brief wanted per-student dashboards; what actually got used was the per-question view, because it told a teacher what to reteach on Monday. We built the dashboards too, and they are the least visited page in the product.
What we would do differently
Feedback generation shipped before we had a good way to control tone, and early drafts read as either clinical or oddly effusive. Teachers rewrote most of them, which defeated the point. Giving each teacher a few tone presets, and letting them save an edited phrasing for reuse, fixed more of it than any model change did.
We would also have built the appeals path earlier. A student disputing an AI-proposed mark is an obvious scenario, and we handled it over email for far longer than we should have.
Questions we get asked
- Does the AI decide a student’s final mark?
- No. The model proposes a mark and its reasoning; a teacher accepts or overrides it, and nothing is published to a student until someone has signed it off. About 94% of proposals are accepted unchanged, but the teacher remains the decision-maker on every one.
- How does it handle written answers differently from numerical ones?
- They go down separate paths. The portal routes by question type first, so a numerical answer is compared against a key with credit for method, while a short written response is judged semantically against a rubric. Treating them identically is what broke the client’s previous attempt at automated grading.
- What happens to a teacher’s corrections?
- Each override is stored alongside the original proposal and the rubric item it related to, then used in tuning. Corrections cluster in practice, and those clusters are the most reliable indicator of which rubric items the model is misreading.
- Can a student challenge a mark?
- Yes. Because every mark carries the model’s reasoning and the name of the teacher who confirmed it, a disputed mark can be traced end to end. Building that path into the product took us longer than it should have.
Have a problem
worth solving?
Tell us what you're building. We'll help you work out what's actually possible, and say so if we're not the right people for it.