KAISAN

This was freelance work with a fixed scope, ending in a handoff. That means the code has to be continued by someone else without me, and no part of it may only work when I am the one running it.

The people using it every day are teachers, not technical operators. Everything has to be manageable through an interface, not through commands on a server.

AI-generated questions are not always correct, and a wrong question on a graded exam damages a student's trust while burying teachers in disputes. So automation must not reach the point of publishing.

There was also a limit I set myself: adaptive difficulty needs a numeric rating, but a visible number immediately becomes something students compare with each other.

I chose Laravel 12 with Livewire 3 and Filament 4. Filament gives a mature admin panel without building management screens from scratch, and that matters on fixed-scope work — the time goes into business rules, not CRUD tables.

Teacher review is a required step before a question can go live. AI output enters as a draft, and there is no path that skips this stage. Automation stops exactly before the point where a student bears the consequence.

Adaptive difficulty uses Elo ratings, but the raw rating is never shown to students. They see a band label only: Pemula, Berkembang, Mahir, Ahli. The system keeps a number to select questions; students get no number to rank each other by.

Scheduled exams and a seasonal leaderboard give the term a rhythm. The leaderboard resets each season, so a student who starts late is not chasing an impossible gap.

PostgreSQL was chosen so the client can run it with any provider after the handoff.