Learned by building.

What I did. The job was tutoring for my daughter: a foundation in the language, an unfamiliar exam, and no way for a parent to see what she actually knew. I chose it because I know the field. I have been a Chief Learning Officer; I understand education technology and what I wished existed. Then I asked whether the need was hers alone. It was not: a private tutoring market in the low billions, parents paying more and seeing less.

Lesson. The need has to be sharper than the technology, and you should build where your judgement is already good. Every AI project I have seen fail started with the tool and went looking for a problem.

What I did. Listed what did not exist in the market and would make the product worth having: a pipeline that turns an exam board’s own specification into a full course; a parent loop that says what to do tonight; personalisation to how one child learns. Then I sorted everything else into “best in class” or “good enough”. Magic-link sign-in only needs to be good. The ingestion pipeline has to be the best there is.

Lesson. Decide where you will be excellent before you build anything. A product that tries to be best in class everywhere is late everywhere.

What I did. Worked out what a family spends on tutoring and schooling, what an AI lesson could cost to serve, what margin that left, and what the product would have to prove before a parent kept paying. Wrote the strategy as three horizons and chose where to win first: one exam board, one subject family, one country, one proof (measured grade improvement in a small cohort). Wrote down what I would not build yet.

Lesson. Value sizing is a design input. The AI cost ceiling I set here became an engineering law, and the engineering got better for it. Strategy is mostly the list of things you will not do this year, written where the builders can see it.

What I did. Talked the product requirements document into existence over several sessions, then had two different models draft and compare it, and iterated until the two agreed. The PRD owns outcomes, not features: a twenty-minute lesson that loads in under two seconds, a coach that hints rather than answers, a parent brief with three specific actions, exam practice that prepares, sits, marks, reviews and remediates. Every requirement measurable. And a short set of principles that outlived everything else: efficacy first, parents as champions, one source of truth per concern, cost as a service-level objective, privacy by design.

Lesson. A requirement that cannot be tested is an opinion. The PRD was rewritten five times in a year; the principles barely moved.

What I did. Before a line of business logic, decided the shape of the system and the rules of the work. Six layers with contracts at every edge; business policy separated from AI plumbing; one store for each kind of truth (documents, relationships, retrieval) and immutable source files as the only authority for anything official; coding standards, strict typing, the test frameworks and the checks that run before code can land. Twenty architecture decisions were written in the first five weeks.

Lesson. The boring decisions compound. A year on, the architecture has not changed, and it is the reason two AI agents could work in the same codebase without wrecking it. Rigid boundaries are what make speed safe. Debt you avoid on day one is debt you never pay.

What I did. Three deep dives, one per killer feature. The ingestion pipeline: take the exam board’s canonical PDF, map it, and turn it into learning sprints, lessons, lesson plans and tailored interactive widgets; done well, that is a full course of around 150 lessons in about half an hour for a few pounds of AI cost, which changes what a curriculum costs to make. The parents’ evening on steroids: I get five minutes with each teacher once or twice a year and walk away with little I can act on, so the product writes the parent a brief that says what changed, why it matters and what to do tonight. And learner DNA: around fifty signals per student, so that the pedagogy, the delivery, the pace and the moment a hint arrives are tuned to how that child learns rather than to the average child.

Lesson. This is where the model earns its place, and where you decide what it may do, what it may only recommend, and who checks. The exam shell, the timer, the mark ledger: code, never a model.

What I did. Epics and backlogs, GitHub, one pull request per change, and from the very first commit the tooling that keeps a codebase honest: type checking, linting, unit tests, browser tests, architecture guards. Two AI coding agents did the typing under written rules; the rules were maintained like code.

Lesson. Put the guardrails in before the speed, not after. The agents are fast; the guardrails are what make fast safe. The next section is about what happens when a guardrail is missing.

What I did. Sat every lesson, sat every mock, watched where the product was wrong before anyone else could. The machine marked one paper at 26 out of 40 where 34 was fair; the learner was right. Nothing in a schema catches that. A person who knows what right looks like does.

Lesson. Automated proof tells you the system did what you asked. Only a human walkthrough tells you whether what you asked was right.

What I did. Shipped it, gave it to friends, family and pupils at Maren’s school, and went to the school more than once to talk to the headmaster and the head of languages. Then listened, and went round again.

Lesson. The loop does not end. Real users are the only evidence that counts, and every round sends you back to step one with a sharper need.