Education & EdTech Software

We set an explicit performance budget based on an entry-level Android device on a throttled connection, and we hold the build to it. In practice that means server-rendered or statically generated pages wherever possible, a small JavaScript payload, compressed and correctly sized images, and no dependency on a large client-side framework bundle to render text a learner needs to read.

It also means the interface degrades sensibly. When the connection drops mid-lesson or mid-attempt, the learner should not lose their work or be dumped back to a login screen. Local autosave, resumable uploads and a clear offline state are worth more to actual completion rates than any amount of interface polish.

The load profile of an education platform is unusually spiky and unusually foreseeable: an entire cohort arrives within a few minutes of a scheduled start. This is a solvable engineering problem, but only if it is planned for rather than discovered.

We load-test against realistic concurrency before launch, push anything heavy — result computation, report generation, notification fan-out — onto a message queue so it never competes with a learner's request, cache aggressively where content is shared, and run on infrastructure that can scale out for a known peak and scale back afterwards. Our Kubernetes and RabbitMQ work is what makes that routine rather than heroic.

The moment a result affects admission, certification or a job application, it will eventually be challenged. A platform that cannot reconstruct exactly which questions a candidate saw, in which order, what they answered, when, and how the score was computed has no answer to that challenge.

So we log attempts as an immutable record rather than a mutable score row: question selection, presentation order, every answer change with a timestamp, timing data and the scoring rule version applied. Question banks support pooling and randomisation so no two candidates receive an identical paper, which reduces leakage value considerably. Where the stakes justify it we integrate a specialist proctoring provider rather than building surveillance features ourselves.

Education platforms hold data about children, and that changes the calculus on every collection decision. We work through the data set field by field and remove anything that cannot be tied to a genuine educational or operational purpose — the strongest protection remains not holding the data at all.

What remains gets least-privilege access designed around real roles, so a subject teacher, a class teacher, an administrator and a parent each see only what their role requires. Retention is enforced by the system rather than by intention, access to identifiable records is logged and queryable, and consent is stored as versioned state that the platform actually honours when it changes.