What I learned building a math fact app for my third grader (and why the 3-second threshold mattered)

yoan ante — Tue, 12 May 2026 15:04:16 +0000

I'm a developer in New Jersey. About a year ago my daughter, who was in third grade, was hitting the wall on basic math facts. She knew the strategies. She could derive an answer. But the recall was slow enough that anything that depended on the fact, word problems, division, fractions, was eating up all her working memory and crashing the rest of the problem.

I tried the apps everyone tries. Xtramath, Reflex, IXL, Prodigy. They all do practice, and they all give you a session percentage at the end. None of them did the thing I actually needed. Some of them I don't even know how they are allow to call themselves educational apps. Looking at you Prodigy.

So I built it.

The mechanic

The engine is a speed adaptive Leitner plus SM-2 hybrid. Each fact lives in a box. Get the fact right in under three seconds, the fact promotes one box and the interval to the next time she sees it widens. Get it right in three to five seconds, the fact stays in the current box. Take longer than five, the fact demotes two boxes. Wrong, reset to box one. Combined with the spaced reintroduction, the slow facts keep cycling back at short intervals and the fast ones move out of the active queue.

Three to five minute sessions. Stop anytime. That part is non negotiable, because the moment a kid feels trapped in math practice the whole thing breaks and you'll never get them back.

What surprised me

The piece I did not expect was the diagnostic value of the three second cutoff. Session percentage hides the difference between a fact a kid actually retrieves and a fact a kid is reconstructing each time, those two things look identical on paper because they both produce the right answer, but they are completely different from a working memory standpoint and the kid who is reconstructing every fact will hit a wall the moment the problem gets multi step.

The threshold separates them. You can see exactly which facts the kid has and which ones she's faking. That's the actual signal. Everything else is noise.

What the dashboard shows

The stats page surfaces the top 10 most missed facts for each learner, with average response time and times seen on each one. Free tier shows the number one missed fact, the upgrade unlocks the rest. The average response time column is the one that surprises parents most. You can have a fact at 80 percent accuracy but with a six second average response time, which means the kid is still reconstructing every time and the fact is going to break under multi step problems. The list orders by a blended weakness score, not raw error count, so what surfaces is the actual cognitive load driver and not just whichever fact got missed most often last week.

What I built it with

Vite plus React 19 plus TypeScript on the frontend. Material UI for components, Framer Motion for animations, Konva for the scene builder canvas. The backend is Firebase: Firestore for state, Firebase Auth for the sign in paths (anonymous, Google, email link, plus a learner profile PIN flow I built for kids who shouldn't have their own email yet), Cloud Functions in Node 22 for fact provisioning and lifecycle stuff, Firebase Storage for the scene assets. Stripe Checkout for billing, Postmark for transactional email.

The scheduler is time driven, not session count driven. That matters because a kid who blasts through a session getting easy facts fast shouldn't be promoting facts she hasn't actually consolidated. Each session is built from due cards first, then active learning cards (box 3 or below), then new unseen cards. Cards that hit box 4 or above during a session drop out of the active queue so they don't keep getting tested while she's tired.

Where it is now

I am not running pilots. The app has no users to speak of yet except for my daughter. My mom is a teacher, my daughter's third grade teacher is the first person I've actually shown it to outside the family. I have a free tier for everyone and a separate offer of full access for any teacher who wnats to try it with a class, because what I need right now is feedback from real teachers, not signups.

If you have a kid in elementary or you teach elementary math, the site is at https://mathbuilders.com and I'd take any feedback you have, even harsh ones, especially harsh ones honestly.