Know exactly where a student stands. Not how they feel about it.
Machine learning measures what a student actually knows, topic by topic. AI explains it to them and to their parents, and it cannot invent a single number.
JAMB preparation for the 2.2 million Nigerian candidates who sit it every year. The first product on the myndtrace platform.
2,243,816
candidates registered for the 2026 UTME. The number renews every year.
Almost all of them prepare the same way. Push forward, and hope.
- No plan. Students open a past-questions book at question one and keep going. Nothing tells them what to work on, in what order, or when to move on.
- No measurement. They cannot see which topics are costing them marks. Effort goes in and lands nowhere in particular.
- No one watching. Parents cannot make the time, and the syllabus has moved past what they remember. The person paying for the preparation is the one who can see the least.
A student can answer three thousand questions, feel like they worked hard, and walk into the hall with no idea what is waiting for them.
Every prep app now claims an AI tutor. Most will invent your score on request.
Ask one how you are doing and it will tell you. Confidently. In a number it made up moments earlier, because there is no measurement underneath it. It is guessing, in a friendly voice.
For an exam that decides where a young person spends the next five years, a system that fakes certainty is worse than no system at all.
This is not another GPT wrapper.
“Ada is at 72% mastery in quadratic equations and should improve by about 8 points next week.”
Emits numbers, and predicts a future score — two rules broken at once.
“Ada has moved from Developing to Building in Algebra. There is not enough data yet in Geometry.”
States a change in level and names a gap. No numbers, no prediction.
Real numbers. And an honest gap where the numbers are not there yet.
Every topic carries a figure: the percentage correct on the questions actually attempted, how that has moved over time, and how much evidence it rests on.
And where a student has not answered enough for the measurement to mean anything, we do not fill the space with a guess. The topic reads not enough data yet, and the next practice session goes and finds out.
Then it tells them what to do next, and why. Not a generic study plan. The specific topic where marks are going, chosen from their own answers.
A weekly read you can trust, in thirty seconds.
Parents get a short summary in plain language, and the figures that matter. Not analytics to interpret. Not a dashboard to learn.
The student approves the link themselves, and parents see a summary rather than every individual answer. That boundary is deliberate. A student who knows they are not being watched over the shoulder keeps using the product honestly, and honest use is the only way the measurement stays true.
Three layers. The boundary between them is enforced in code.
Measure
Every answer is written once, permanently, tagged to a topic. For each topic we hold an estimate, an uncertainty, the evidence behind it, and how recently it was practised. The model works from the very first answer and states its own confidence, so a topic with three answers is never shown as though it had thirty.
Decide
A weighted function runs across every topic, balancing how much it is needed, how confident we are, how recently it was seen, and how much of the syllabus is covered. Where the evidence is thin it broadens instead of asserting a weakness it cannot support.
Communicate
Only now does language get involved. The model receives a closed list of what it is allowed to mention, and a validator discards any sentence containing a number or naming anything not on that list. Every figure on the screen is drawn by the interface from measured state, never written by a model.
And it is built to improve. Events are ground truth and are never overwritten, so learner state is recomputed rather than stored. When the model gets better, it replays across every student's entire history. Nothing has to be thrown away to improve it.
One engine. Every exam that matters.
The measurement layer does not care what is being learned. Each product inherits it, and every session on any product improves the measurement for all of them.
- JAMB prep Nigeria’s university entrance exam. English, Mathematics and Biology at launch. In development
- WAEC prep Largely the same candidate, sat months apart. The corpus changes; nothing underneath it does. Planned
- NECO prep The third leg of the Nigerian secondary examination path. Planned
- SAT prep The same engine aimed at students applying abroad. Planned
- Office prep Staff training for organisations that need evidence the training actually landed. Exploring
The three things this needs, in the same people.
- Six years teaching this exact exam. Online and in person, with the students and parents this product serves. The idea did not come from a market report. It came from sitting across the table and being asked, over and over, whether it was working.
- Machine learning in production. Shipped paid AI projects, and delivered AI training to organisations including the Federal Competition and Consumer Protection Commission and the National Centre for Artificial Intelligence and Robotics.
- Built AI-native from the start. Not a conventional team retrofitting AI onto an existing product.
Two further core members join the team after version one ships.
Be in the first cohort.
We are opening JAMB prep to a small first group of students and parents. Small deliberately: we would rather a hundred people trust what we tell them than a hundred thousand ignore it.
You’re on the list
We’ll be in touch when JAMB prep opens. Thank you.