The First Version of AcademiaPlus Failed

I like telling this story because it is not flattering. AcademiaPlus, as it exists today, only exists because the first version did not work.

The first idea

The first AcademiaPlus was a fairly conventional idea: use AI to generate assessments. A teacher describes what they want, and the system produces a curriculum outline, then questions, then marking. It sounded efficient. It was also designed to work for anyone, anywhere, which should have been a warning.

What real teachers told me

The trouble started as soon as I put it in front of real teachers.

The AI-generated curriculum did not hold up next to a teacher with fifteen years of experience, who knew exactly how the real curriculum is structured and taught. The generated questions felt foreign next to real Rwandan exam papers, which have their own structure, their own way of building multi-part questions, their own rhythm. A platform built to work for everyone ended up working authentically for no one in particular.

Nobody needed to say it bluntly. It was a general tool with Rwanda's name on it.

What I changed

Instead of adjusting the tool, I changed the foundation. AcademiaPlus is now built around what is actually specific to Rwanda.

• The national curriculum, built in. Every REB and TVET curriculum and syllabus structure is loaded once and shared by every school. Nobody uploads a curriculum or asks an AI to invent one.
• Real questions. A growing repository of human-authored exam questions, drawn from national and school past papers and contributed by teachers, each one mapped to a precise point in the curriculum. AI-generated questions are meant to be a small minority, not the default.
• Real exam structure. Questions with several parts, shared context, diagrams and different answer types, including math working, graphs and diagrams, so a student can answer the way they would on a real paper.
• Rwanda's real academic rhythm. The national calendar, terms, promotion, repeating and transfers, so a school that joins mid-year still gets a complete and correct academic year.

Where AI fits

AcademiaPlus is not an AI assessment platform, and I push back when it gets described that way. AI plays a supporting role: it helps with grading, it digitizes past papers into structured questions (always with teacher review before anything is saved), and it fills gaps in coverage. The foundation is the curriculum, the content and the evidence.

Because every question maps to a specific point in the curriculum, every graded answer becomes evidence about how well that point is being learned, by one student, a class, or eventually a whole school. That is what turns AcademiaPlus from a tool for making tests into a way of seeing what needs to be taught next.

What I still have to prove

I don't want to oversell this. The rebuilt version is still being finished and tested, and the real test is not my confidence but whether teachers and students use it and trust it.

The plan is deliberately small: put it in front of a first school, see what breaks, and expand only when the evidence says to. I am also not guessing prices before real usage exists. Guessing early was one of the mistakes of the first version, and I would rather learn what it really costs and what it is really worth from actual schools.

The lesson

Building for everyone is a comfortable way to build for no one. The version that failed taught me to start from the constraints, the real curriculum, the real exam papers, a real teacher's judgement, and to treat them as the product instead of obstacles to it.