Personalisation starts with listening, not an algorithm.

User Research

The personalisation work began with a recurring learner problem: progress was hard to feel, and the next useful thing to study was hard to choose. Research gradually turned that into a product model: respond to what just happened, show what you can now do, then use that history to decide what should come next.

Role
Principal Designer
Context
Personalisation
Scope
Usability, Behavioural Data
Years
2020 — 2022

Research mattered when it changed what the learner saw next.

The recurring problem was not a shortage of content. Learners struggled to tell what needed work, whether they were improving, and what was worth studying next.

Interviews exposed those gaps; prototypes made the choices testable. Learning history gradually gave the product enough context to respond differently.

Ten CEFR-J levels — Progress reframed as concrete speaking capabilities

One daily plan — Learning history shaping what appears next

AI Coach (2020)

Turn feedback into something the learner could act on next.

AI Coach was an early attempt to make assessment useful beyond the score itself. Free-answer tasks were broken into subtasks and marked pass or fail individually, and feedback pointed back toward what needed work and what to practise next — creating a simple loop between performance and the next action.

CanDo Progression (2021)

Make progress look like something you can now do.

Learners could not tell whether they were improving. CanDo reframed progression around CEFR-J-based speaking capabilities across ten levels: each lesson had something concrete to work toward, while accumulated Can-dos made longer-term progress visible.

Buddy Tab (2022)

Turn the home screen from a curriculum list into today’s work.

Buddy Page brought the next decision to the front of the product. Instead of opening onto a curriculum index, learners were greeted by a Buddy and a finite daily plan shaped around their study settings and progress. A callout named that day’s personalised exercise by name before the learner even opened it. The first screen could now answer a simpler question: what should I do today?

Personalized Exercises (2022)

Use learning history to change what practice appears next.

The schedule became more useful once the contents could react to the learner. Learning logs could expose weak areas and surface personalised missions rather than treating every learner as the same queue. The slot itself rotated between three distinct formats — vocabulary recall, phrase drilling, shadowing — each reading a different kind of weakness. The question shifted from what content exists to what deserves the learner’s next few minutes.

KeyPhraseDojo (2022)

Drill the exact phrase, then prove it landed out loud.

KeyPhraseDojo took a single lesson’s key phrase and turned it into a short, repeatable loop: pick the correct phrase out of near-miss distractors, then say it aloud. Spoken attempts were scored phrase-by-phrase — native-like or not clear — so a three-minute drill could sit inside the rotating personalised-exercise slot alongside VocabTherapy and CopyCat.

CopyCat (2022)

Shadow a native speaker, see exactly where you drifted.

CopyCat had the learner repeat a line after an AI voice, then laid their intonation curve over the target’s, word by word, with phonetic transcription underneath. Tempo was scored separately from pitch against an ideal pacing window, so “too slow” and “wrong shape” became two different, actionable notes instead of one vague score.

VocabTherapy (2022)

A prototype for deciding what should return before it was forgotten.

VocabTherapy pushed the same idea into memory. Listening, word and phrase recall were treated as different signals so review did not have to mean repeating everything. The prototype disappeared, but the problem did not: later features would use individual learning history to decide what should come back and when.

I can’t tell whether I’m improving or standing still. — Translated user feedback cited in Can-do launch, 2021

Make the adaptation visible enough to trust.

The system could become increasingly sophisticated underneath and still feel arbitrary. Each personalised surface had to make the next action understandable:

Research turned personalisation into a visible product model.

AI Coach connected performance to the next action. CanDo made longer-term progress understandable. Buddy Page made the day’s plan visible, and personalised practice let learning history influence what appeared next. The common move was not more AI; it was reducing the number of learning decisions the learner had to make alone.