AI is most useful in education when it is part of the learning loop, not a button added after the product is finished. I work on an English-learning system for Chinese learners across web, Android and WeChat. The product combines structured course material with vocabulary, reading, listening and guided practice.

Start with the learning loop

The core loop is deliberately simple:

assess → select material → practise → explain → review → adapt

Each step has a clear job. Assessment estimates the learner’s current level. Material selection keeps vocabulary and sentence structure appropriate. Practice produces observable work. Explanation makes errors understandable. Review converts those errors into the next exercise.

AI can support every step, but it should not own the entire loop. Level rules, curriculum structure and safety constraints remain deterministic. Models operate inside those boundaries.

One product, several clients

The same learning experience must work across a browser, a WeChat mini program and native mobile applications. The difficult part is not rendering the same screen everywhere. It is keeping product behaviour consistent while respecting each platform.

I treat the existing product as the behavioural contract:

  • the same copy and learning sequence;
  • the same loading, error and empty states;
  • the same account and progress model;
  • thin platform adapters for recording, storage, payments, updates and sharing.

Shared business rules are valuable. Shared UI code is valuable only when it does not make every platform feel like a web page inside a wrapper.

Production is part of product design

Educational software is a system, not a set of screens. Content publishing, model configuration, database migrations, Android builds, mini-program releases and rollback paths all affect the learner’s experience.

The engineering goal is therefore not maximum novelty. It is a product loop that is understandable, measurable and safe to change.