MathsTutor is our flagship and our teacher. Building it is where the thesis on the rest of this site stopped being theory. Some of what we believed going in survived contact with the work; some of it didn't. This is the honest version.
1. We built a chatbot. That was the wrong centre.
The obvious shape for an AI tutor in this era is a conversation: the student asks, the tutor responds, Socratically, guiding rather than telling. We built toward that. Then the evidence got in the way. Open-ended chat, it turns out, is most of what students use to cheat — and as a primary teaching mode it's weaker than the thing it replaced. The stronger pattern is structured: a generated lesson, a worked example, then practice, with the conversation reserved for checking understanding, not carrying it.
That meant rethinking what a "session" even was. It was the most useful wrong turn we took, because it forced the question every EdTech builder should ask first: not "how do we use AI," but "how do children actually learn," and only then "where does AI help."
2. Eighty per cent isn't mastery
Our early bar was the industry default: pass a topic at around 80% and move on. The Alpha School evidence reframed it bluntly — a student at 80% is still losing the ball one time in five, and those misses compound into the gap that shows up at the exam. Mastery means near-total recall, demonstrated across more than one session, not a single passing score on the day. We moved the gate from "completed the lesson" to "can still do it cold, later." It's a harder bar, and it's the point.
3. The knowledge graph is the engine, not a feature
When a student fails at quadratics, the naive system drills quadratics harder. The right system asks a different question: what's actually missing? Often it's factoring, or negative numbers, from two years earlier. The mechanism that makes mastery learning work is a map of prerequisites — detect the real hole, drop the student there quietly, fill it, then bring them back. We had listed this as a capability. Building it taught us it isn't a capability; it's the primary logic. Everything else hangs off it.
4. Engagement is the wrong metric
Every instinct from consumer software says: maximise time in the app. Education inverts it. The evidence is that a focused 25-to-35-minute session with a hard "you're done for today" ending beats an open-ended grind, and that long sessions cause disengagement even when the content is good. So the product shouldn't reward a student for staying; it should celebrate finishing and stop. A child who masters the day's work in twenty-two minutes and logs off has won. Building that meant deliberately not shipping the engagement features that would have made the dashboards look better.
5. Parents don't buy "an AI tutor"
We learned this the way everyone does: by writing copy that didn't land. Parents don't buy "AI-powered personalised maths tutor." They buy an outcome with a clock on it — "caught up before the exams," "from one grade to the next in a term." The thing that unlocks the decision is time and a grade, not the elegance of the technology. It changed how we talk about the product entirely: lead with how fast and by when, and keep the technology where it belongs, underneath.
6. The moat is teaching, not answering
The tools students already use — general chatbots, photo-solvers, computational engines — all do the same thing: they give the answer. A tutor that teaches instead of answering is doing the opposite job, and that difference is the whole position. Parents are already uneasy about their children shortcutting homework with AI; a product that visibly refuses to be a shortcut is solving a fear, not just a maths problem. "The AI that teaches, not the AI that cheats" wasn't a slogan we started with. It's one the build handed us.
What we haven't proven yet
Here's the part most case studies leave out. MathsTutor is a real, working product, and everything above is what building it taught us. What it is not yet is proven on outcomes. The claim that matters — real before-and-after scores from real students, showing the gap closing — is the thing we still have to earn, and we won't dress up a roadmap as a result. That honesty is a commitment, not modesty: the day we have the number, we'll publish it, and until then we say so plainly. Anything else would fail our own test.
Frequently asked questions
Is MathsTutor live? It's a real, working product and our flagship build. What we're still earning is published outcome data — before-and-after scores — which we won't claim until we have it.
Why did you move away from a chatbot? Because open-ended chat is both the main way students cheat and a weaker primary teaching mode than structured lessons. We kept conversation for checking understanding, not for carrying the teaching.
What's the most important thing building it taught you? That the right first question is "how do children learn," not "how do we use AI." Get that order wrong and you build a clever product that doesn't teach.
A note on why we wrote this. Addestra co-builds EdTech ventures from zero to one. MathsTutor is our own build, and publishing what it taught us — including what we got wrong — is how we'd want any partner to work with us. How we work: editorial standards.
Sources
- Education Next — "Two-Sigma Tutoring: Separating Science Fiction from Science Fact" (the mastery-tutoring evidence base): https://www.educationnext.org/two-sigma-tutoring-separating-science-fiction-from-science-fact/
- Our framework: System-Native EdTech, The Four Mismatches Test