From reading to math, science, and languages: the Lumikids roadmap

The first version of Lumikids only does one thing. It listens to a four-year-old sound out a word and answers him in under a second. That's it. Not a phonics curriculum, not a leveled library, not a math module, not a science quiz. One subject, one voice loop, one kid at a time.

People who hear the pitch usually ask the same two questions. Why reading first? And when does math show up?

This post is the honest answer to both.

Why we started with reading

I started with reading because it is the highest-stakes subject we hand to small children, the one with the most settled science, and — paradoxically — the one that almost every Artificial Intelligence (AI) tutor on the market does worst.

The stakes are obvious to any parent who has watched a third grader fall behind. A child who is not reading fluently by the end of second grade is statistically very likely to struggle across every other subject for the rest of their schooling, because every other subject eventually arrives as text. Reading is the load-bearing wall. If it cracks early, everything above it settles wrong.

The science is also more settled than the public debate suggests. The "reading wars" are mostly over in the research literature; structured, phonics-based instruction works for the overwhelming majority of kids, and the rest benefit from it plus targeted support. Organizations like the National Center on Improving Literacy and the work cataloged by the What Works Clearinghouse have spent decades narrowing down what actually moves the needle. We didn't have to invent a pedagogy. We had to build a delivery mechanism that didn't waste it.

And then there's the technical reason. Reading is the hardest thing to do well in voice-first AI. A child sounding out "frog" produces audio that bears almost no resemblance to an adult saying "frog." The tutor has to recognize the attempt, judge whether it was close enough, decide what to say back, and say it before the kid loses the thread. If you can build a system that handles a four-year-old's mispronounced phonemes in under a second, math is, frankly, easier.

I wrote about the moment that started all of this in how my four-year-old taught me to build an AI tutor. Reading was where the problem was loudest. So reading was where we started.

Math: conversational problem-solving, not flashcards

The math version of Lumikids will not be flashcards.

There are already a hundred good flashcard apps. Some of them, like the early modules of Khan Academy Kids, are genuinely useful for drill. We are not going to build a worse version of those. The thing missing from kids' math software is not more drill. It's the conversation that a good math tutor has with a kid who is stuck.

What that looks like in practice: a six-year-old says "seven plus five is eleven." Most apps mark it wrong, show a sad sound, and move on. A real tutor pauses and asks, "How did you get there?" Sometimes the kid says "I counted seven and then four more," and now you know exactly which finger they lost. Sometimes the kid says "I just guessed," and now you know to slow down. The diagnostic move is the teaching.

Lumikids math will work that way because the voice stack already supports it. The same loop that lets Remi argue with the tutor about whether he really said "fr-oh-g" or "frog" will let a slightly older kid talk through why they think a half of eight is three. The tutor's job is to ask the next useful question, not to grade.

We'll start with the operations and concepts kids encounter from kindergarten through about second grade — counting, place value, the four basic operations, fractions as fair-sharing, simple word problems. That's enough to keep us busy. We are explicitly not chasing standardized test alignment.

Science: curiosity-driven, not curriculum-driven

Science is where Lumikids stops looking like school software at all.

Kate and I have a rule about Remi's questions: when he asks one, you stop what you're doing and follow it. When he wanted to know why the moon is sometimes out in the daytime, we went outside and looked at it. When he wanted to know how a zipper works, we took one apart. That rule was the seed of the whole product, and science is where it gets to grow.

The science version of Lumikids will not have a unit on weather followed by a unit on plants followed by a quiz. It will have a child saying "why is the sky pink right now," and a tutor that can talk about Rayleigh scattering at a four-year-old's level for about ninety seconds, then ask the child what color they think the sky will be in an hour. The "curriculum" is whatever the kid is currently looking at.

This sounds soft until you remember that every working scientist started as a child who would not stop asking why. The thing we are trying to protect is not facts about photosynthesis; we can hand a kid facts at any point. The thing we are trying to protect is the asking. Most kids' science apps train the opposite muscle — they reward correct multiple-choice answers about science instead of the act of being curious about it. We'd rather build the curiosity engine and let the facts accumulate behind it.

We'll lean on primary sources for the underlying content. NASA's Space Place and the Smithsonian's open science resources are the kind of references we want the tutor to be able to draw on without making things up. When the tutor doesn't know, it says so. That's a feature.

Languages: the voice stack is already built for this

Languages are the most obvious extension and probably the easiest engineering lift, because the entire product is already a voice loop. Built on Anthropic's Claude for reasoning, ElevenLabs for sub-second voice synthesis, and Wispr Flow for understanding the messy real speech that comes out of small humans, Lumikids is, in a sense, already a language tutor. It just happens to speak only English right now.

The interesting design question for languages is not technical. It's pedagogical. Do you teach a five-year-old Spanish by having a tutor switch entirely into Spanish and trust immersion? Do you scaffold in English and gradually fade it out? Do you let the kid pick the topic and translate it on the fly? We have opinions but not certainty, and we'd rather pilot it with a small group of families than guess.

I'd expect Spanish first, then Mandarin, then French — driven by the languages we have native-speaker testers for, not by market size. The voice work we did to support voice-first learning around speech, not taps means we don't have to rebuild the foundation. We just have to teach the foundation new words.

A realistic timeline

I'm not going to give you dates I can't promise.

Reading beta is live now. Math is the next subject we're prototyping, and I'd rather ship a small, narrow math experience well than a wide one badly. Science and languages come after that, in whatever order our beta families and our hiring tell us makes sense.

What I can promise is the order of operations. We finish reading — meaning it works reliably for a range of ages, learning differences, and accents — before we widen the surface area. Every subject we add has to clear the same bar reading is being held to: sub-second response, parent-visible session data, no engagement-maximization patterns, no data sold or shared. If a subject can't ship at that quality, it doesn't ship.

What we're not going to build

Two products keep getting suggested to us. We're not building either.

No test prep. There is a healthy market in tutoring kids for state assessments, the Scholastic Aptitude Test (SAT), and the rest of the alphabet. It is not the market we are in. Optimizing a child for a specific test is the opposite of protecting their curiosity, and we'd rather be a smaller company than a bigger one that does this.

No homework helper. Lumikids will not be the app a fifth grader opens to get their math worksheet done at nine p.m. We have nothing against homework help as a category, and tools like Khanmigo are doing thoughtful work in that space for older kids. But a homework helper is structurally a tool for finishing assignments, and the whole point of Lumikids is to be a tool for starting questions. Different product, different incentive, different company.

The honest version

The roadmap above is what we believe today. Some of it will be wrong. Math may turn out to need a different voice tuning than reading. Science may need a feature we haven't thought of yet to handle a kid who wants to talk about dinosaurs for three weeks straight. Languages may go in an order we don't expect.

What won't change is the constraint. One kid at a time, in a voice loop fast enough to stay inside their attention, with parents who can see exactly what happened and a company that doesn't make money by keeping their child on the screen longer. Every subject after reading has to fit that shape, or it doesn't get built.

If you want to be in the room while we figure out which parts of this plan survive contact with real children, join the beta.

Image brief

• Hero image: A child's hand drawing a winding path on butcher paper from a small open book labeled "reading" to hand-drawn icons for math, a leaf, and a globe, warm overhead light, crayons scattered.
• Inline image 1: A simple roadmap diagram with four stops — Reading (lit up), Math (next), Science (dimmed), Languages (dimmed) — placed after the "Why we started with reading" section.
• Inline image 2: A side-by-side sketch of a flashcard app versus a conversation bubble showing a tutor asking "how did you get there?" — placed inside the math section.

Internal link suggestions

• "How my four-year-old taught me to build an AI tutor" — anchor: "how my four-year-old taught me to build an AI tutor"
• "Voice-first learning: why we built around speech, not taps" — anchor: "voice-first learning around speech, not taps"
• "Adaptive learning isn't a setting — it's the whole product" — anchor: "how Lumikids actually adapts to each kid"

Editor's note

Tim to confirm: (1) the order of language rollout (Spanish → Mandarin → French) and whether we actually have native-speaker testers lined up, or whether that should be softened to "the languages our early families speak at home"; (2) whether we want to name Khanmigo by name in the "no homework helper" section, or pull the link and keep it generic; (3) the K–2 math scope I described — confirm we're comfortable committing publicly to that range, or whether to widen/narrow before publish.