← 02 MAKEWhat was actually built?
An ML teaching tool aimed at teachers, not students
A teaching tool that lets you touch gradient descent in real time. The target got re-cut twice — to teachers, not students.
02
Constraint
It had to open in a classroom from a link alone, with no login and no install. That constraint decided the technology.
Given up
A Flutter prototype was built for comparison and then dropped. It suits native distribution, but a no-install classroom tool would still owe a full rewrite.
What remains
Two prototypes implementing gradient descent from scratch with no external ML library. Not submitted yet, and the largest gap is not a feature but the classroom script.
The PRD went from v0.1 to v0.3 in three revisions, and two of them were not about features. They were about re-cutting the target. §3 and §4 were rewritten whole.
The first target was "people for whom code is the barrier." That is weak. Bind people by motivation and the reason to use the thing becomes "because I'm interested" — and interest does not get a tool adopted.
Cut sharply, the answer is this: a teacher who needs a demonstration in the room, now. Teachers do not pick a tool by feature count. They pick by whether friction and risk go down. Students are not the primary user; they are the beneficiary.
Gradient descent, over- and underfitting, the bias-variance tradeoff — the intuitions being taught are unchanged. What changed is whose hands they are put into.
Decided and closed
- Primary user = teachers, students are beneficiaries
- No login, no install — running in a classroom from a link alone is the core constraint
- A demo/projector mode is a must-have
- Student management is deferred to a later phase
- Monetisation comes last — competition entry and portfolio value come first
- Technical path: Next.js + PWA is more practical for a link-only, no-install classroom tool (Flutter was prototyped for comparison only)
What exists now
- The PRD — vision, personas, differentiation, MVP scope, technical architecture, phases. Three revisions from v0.1 to v0.3, with §3 and §4 heavily rewritten while the target definition was challenged twice.
- Two prototypes, both implementing gradient descent from scratch with no external ML library: real-time visualisation, draggable data points, a loss curve, and a rule-based Korean state explainer.
- React/JSX on Next.js
- Flutter/Dart, for comparison
Still under review
Technical depth past the basic visualisation is the condition for competing. Candidates:
- Training a model inside the browser (TensorFlow.js)
- Misconception diagnosis from learning logs
- Replacing the rule-based Korean explainer with dynamic feedback from the Claude API
What it taught
- The Flutter/Next.js tradeoff is real. Flutter suits native distribution but needs a full rewrite, and a backend proxy for the API is still owed. For a no-install, open-from-a-link classroom tool, Next.js + PWA was the practical answer.
- The most valuable thing missing is not a feature. It is the classroom script.