Our Story
Si Math AI began with a pattern that experienced teachers recognise immediately: the same students, making the same mistakes, for the same reasons — and no practical way to catch it in time.
The people who built this platform spent years teaching mathematics to American Diploma students preparing for the SAT, the ACT and the EST. Teaching at that level produces a very particular kind of knowledge. You learn that a student who fails a quadratics question has usually not failed at quadratics — they have failed at factoring, or at sign handling, or at translating a word problem into an equation. You learn which wrong answer choice a student picks when they have made each specific error, and what that choice tells you. You learn that the gap between a 600 and a 700 is rarely more content; it is usually four or five specific, identifiable, fixable skills.
The frustration was never the knowledge. It was the arithmetic of applying it. One teacher, thirty students, two hours a week. You can diagnose a student properly — but only if you sit with them, watch them work, and ask the right questions. There is no version of a classroom timetable where that happens for everyone who needs it. The students who could afford private tutoring got the diagnosis. The rest got told to do more practice questions, which is advice that helps least the students who need help most, because practising what you already know is comfortable and practising what you don't is not.
Then large language models became good enough at mathematics to matter. The first instinct — the one most products acted on — was to point a general chatbot at exam questions and call it a tutor. We tried it. It answered questions well and taught almost nothing. It had no idea which exam a student was sitting. It forgot every mistake the moment the conversation ended. It could not tell a student what to study next, because it had never been told what "next" means for someone eleven weeks from an SAT with a weakness in coordinate geometry. It gave a correct answer and left the student exactly where they started.
That was the insight the platform was built on: the AI was never the missing piece. The teaching was. A model can explain a step. It cannot, on its own, know that this particular student's real problem is sign errors under time pressure, that the fix is a specific drill sequence, and that the drill should come before the next mock exam rather than after it. That knowledge lives in teachers.
So the work was inverted. Instead of asking what an AI could do with mathematics questions, we asked what an experienced SAT, ACT and EST mathematics teacher does across a full preparation cycle — diagnose, explain, assign targeted practice, re-test, re-diagnose, keep the student motivated — and then built the systems required to carry each of those steps. Educators and exam specialists defined the taxonomy of skills, the mistake patterns, the explanation methods and the strategy content. Engineers built the diagnostic pipeline, the mastery model, the mock exam scoring, the memory layer and the mentor that delivers it.
Zero — the platform's AI mentor — came last, and deliberately so. Zero is the interface to that body of knowledge, not the source of it. That ordering is the single most important design decision in the product, and it is why Si Math AI behaves like a teacher who knows you rather than a search engine that answers you.