Roadmap
Where OpenMath is going
A real product, growing in the open. What’s shipped, what’s being built, and where it’s headed.
Shipped
- 24 quiz types across 6 topics
- Badges, streaks & progress tracking
- Real-time multiplayer
- Hungarian & English throughout
- Student, teacher & parent roles
Building now
- More — and harder — quizzes
- A classroom pilot with real 3rd-graders
Up next
- AI-assisted adaptive difficulty
- In-app observability (traces & metrics)
On the horizon
- Migration to Kubernetes
- On-prem AI model serving (private & sovereign)
- A mobile app for iOS & Android
The roadmap is shaped by real classroom use — and built in public. Curious how the production side works? See how it’s built →
AI assist
The three things we want AI to do next
All three run on ATAILA’s own hardware. Nothing a child types is sent to a public AI service, and none of them needs data we do not already keep.
- 1
Explain the mistake, not just the answer
Right now a wrong answer gets you “Wrong — the correct answer is 56”, and that is all. The app already knows far more: which question it was, what you actually typed, how old you are, and what you have been practising. That is enough for one sentence naming the kind of mistake — an off-by-one, two swapped digits, the classic 7×8 and 6×8 mix-up — in your own language, right when it would help.
First, because it is the smallest change with the largest effect, and a parent notices it immediately.
- 2
Practise what you keep getting wrong
Every answer you have ever given is stored with the fact behind it, so the app can already see that the 8-times table is slow and the 2-times table is not. It simply does not act on it — the difficulty is whichever button you pressed. Next, the mix leans quietly toward the facts you actually miss, without turning practice into a punishment.
Worth saying plainly: this one is mostly arithmetic over a table, not a language model. Choosing the right tool matters more than choosing the biggest one.
- 3
Feedback a teacher can send after one edit
Today a teacher reviewing a quiz picks from a handful of stock sentences. Instead, the app drafts one that names what actually happened — “seven of the eight misses were in the 8-times table, and each took over twelve seconds” — and the teacher edits it before it goes anywhere. A model never writes to a child unreviewed.
Third because it saves the most time per use, and a human is in the loop by construction.
And one we are deliberately not doing
We will not let a model invent the sums. The questions come from ordinary code that is correct by construction, and a model that makes up arithmetic will eventually make it up wrong. A machine can help explain a mistake; it does not get to decide what 7×8 is.
One more is waiting on something we do not own yet: reading the question aloud. For a seven-year-old who reads slowly, a timer measures reading rather than maths — but every speech model we run today turns speech into text, not the other way round. That one needs a Hungarian voice first.
Have an idea for a quiz?
New, harder quizzes are added all the time. Tell us what your kids need next.
Get in touch →