For most of my career I have looked at products the way the person building them does. Last year I started looking at one category the way the person paying for it does, because the users were my own kids. What I saw in the learning apps on our tablet has stayed with me, and it is the reason I have spent the past several months building something I was sure, for a long time, that I would never need to build myself.
This post is the first in a series about that build. Before any architecture, I want to write down what I actually saw, because the frustrations came first and every technical decision since has been an answer to one of them.
What the marketing says and what the tablet shows
If you read the landing pages, this is a golden age of digital learning. Adaptive paths, AI tutors, personalized everything. If you sit next to a child using these products for a few weeks, a different picture forms.
The first thing I noticed was how often failure is the loudest moment in the room. A wrong answer lands as a red mark, a falling streak, a sad sound. My kids did not respond to that by trying harder. They responded by avoiding the topics that produced it. The products call this gamification. Watched closely, it works more like a small punishment economy, and the currency is a child's willingness to attempt something they are not yet good at. Learning lives entirely inside that willingness.
The second thing was the typing. Most of these products assume a learner who can read a question silently, hold their reasoning in their head, and type an answer. That is not how children think. Mine talk their way through a problem, out loud, in fragments, half in one language and half in another. A text box throws all of that away. The thinking is the interesting part, and the interface cannot hear it.
The third thing was subtler and took longer to name. A word problem would ask about dollars and dimes, or about names and foods chosen from one place in the world, and my kids would stall before the math even started. They were not stuck on the arithmetic. They were translating someone else's everyday into their own. We have lived between two countries, so I may notice this one more than most parents do, but the cost is real. A child spending effort decoding the context has less left for the concept.
The obvious shortcut, and why I could not take it
By last year the obvious answer to all of this was supposed to be a chatbot. The general models are patient, conversational, happy to explain fractions forty different ways. I tried this, supervised, and the experience is genuinely impressive right up until it is not.
The model would explain beautifully, then confidently get an answer wrong. It would drift into vocabulary no seven-year-old should be decoding. Sometimes it would simply do the work for them, which a child learns to exploit faster than any adult expects. None of this is surprising if you know how these systems work. A language model predicts plausible text. It does not know when it is wrong, and a young learner is about the worst-equipped audience there is to catch it.
I want to be careful here, because I am not claiming chatbots have no place in learning. For an adult who can challenge an answer, they are a remarkable tool. My claim is narrower. An unsupervised young learner and a system that is confidently wrong some unknown percentage of the time is a pairing I am not willing to accept, as a parent first and a designer second.
Trust in a learning tool is not a feature you can patch in later. It is the ground the whole thing stands on.
Deciding to build
For months my position was that someone else would surely fix this. The market is enormous, the problems are visible to any parent paying attention, and the technology has come far enough to make it feel within reach. I kept waiting, and the new products kept arriving as thin wrappers around the same general models, with the same failure modes and a fresh coat of mascot.
At some point the question flipped. I have spent two decades designing and building products, the last stretch of it on services where being wrong has real consequences for real people. My kids are sitting at the kitchen table with the exact problem my whole career has been preparing me to take seriously. Waiting started to feel stranger to me as an option.
So I am building it. A learning engine, math first because that is what my kids need now and because math is demanding enough to keep the design honest. I am deliberately not building a chatbot with worksheets attached. The shape of it is different, and the next post in this series walks through the one decision that everything else hangs on, where the AI sits and, more importantly, where it never sits.
What this series is
I intend to write this as a build log, decisions as I make them, including the ones that turn out wrong. I have read enough retrospectives that smooth the path after the fact. The honest version is more useful, to me at least, and possibly to anyone else building for learners right now.
Two promises for this series, then. Every number I share will be a number my system actually produced. And when I change my mind, the post saying so will name the earlier post it corrects. A build log that cannot admit mistakes is just marketing with timestamps.
What I saw on that tablet was not a technology problem. The technology is more capable than it has ever been, and still moving. It is a priorities problem, products tuned for engagement metrics and parental guilt instead of the slow, fragile, human business of a child becoming confident at something hard. I suspect the fix has to come from people who are watching that fragility up close. I happen to live with the most honest informants I could ask for.