LexAI

Adaptive software does its best work invisibly, which is the problem. I made the personalisation something the learner could see.

Client: Competition project
Role: Product Design
Dates: 2024
Scope: Research, personas, flow and IA, design system, high-fidelity screens, dark mode
Team: Three-person entry with Hanwill and Hady Wijaya
Outcome: 1st Place, Mobile App Design. Coris Indo National, Aug 2024

Context

LexAI adapts lessons to how each learner learns, adds practice with native speakers, and returns feedback in real time. The competition brief asked for an AI language tutor. The design problem underneath it was that adaptive systems do their most valuable work invisibly.

Indonesian press coverage of low English proficiency — The brief was a competition prompt. The problem was already in the press.

The mark.

About the project: an AI-powered language learning app — An AI tutor that adapts, and puts practice with native speakers inside the app.

Problem statement and solution — Existing apps are static. The gap is feedback that arrives while you are still wrong.

Process

The scope of work: five design phases — Five phases, and what each one had to produce.

Empathise: selected statements from user interviews — Empathise: what learners asked for, in their own words.

Define: the four problems — Define: four failures, named rather than implied.

Personas: Nadia, 24, and Carlos, 27 — Nadia wants range. Carlos wants three languages inside two years.

User flow and information architecture — Every screen placed before any screen was drawn.

Colours and typography — Manrope, one blue, one orange — small enough to hold in your head.

Greyscale wireframes — Grey first, so structure could still move for free.

Onboarding wireframes and first prototypes — Onboarding asks two things: which language, and what you care about.

Decisions · 01

Design for the gap between sessions

Chose

Progress tracking and goal setting aimed at the days when the app is closed.

Rejected

Putting that effort into a richer in-lesson experience.

Why

Language learners don't quit inside a lesson. They quit between them, so that's where the design had to work.

Evidence

The streak screen defines itself on the spot — a streak counts how many days you have practised in a row — and the size of the commitment is the learner's to set at 5, 10, 15, 30 or 60 minutes before day one.

Study-time selection and the streak screen — Five minutes to an hour — the learner sets the commitment, then the app holds them to it.

The home screen — Today's challenge at 75%, the AI tools under it, practice under that.

Daily Challenge and Leaderboard — Four tests a day at your level, and a leaderboard that gives the streak somewhere to go.

Decisions · 02

Show the learner what the system changed, and why

Chose

Surfacing each adaptation in plain language at the moment it happens.

Rejected

Silent adaptation, the default and the smoother interaction.

Why

If the app quietly adjusts your lesson and never says so, it reads as an ordinary app that changed its mind. Made visible, the same behaviour reads as attentiveness.

Evidence

The Smart Reading Tutor returns pronunciation as one chip per word rather than a score, and the Speaking Test transcript marks each error at the point in the sentence where it happened.

The Speaking Test: recording, transcript and results — Your own recording, transcribed, with each error marked where it happened.

The Listening Test — Listening, answered against audio you can replay rather than a memory of it.

The Reading Test — Reading, scored against the passage still on screen.

The Writing Test — Writing, with the brief kept on screen while you write to it.

Smart Reading Tutor with per-word pronunciation correction — The tutor names the words you missed, one chip each, instead of scoring the sentence and moving on.

Brainy Room Chat — A room for the open-ended asks that never fit inside a lesson.

Decisions · 03

Break practice down to the unit that actually fails

Chose

Practice split to the individual grammar unit — independent, dependent, noun, adjective and adverbial clauses — each with its own fifteen questions and its own score.

Rejected

One Grammar module with a single progress bar.

Why

A learner who keeps missing adverbial clauses does not need more grammar. They need that one unit. A combined score can only tell them to try harder, which is not an instruction.

Evidence

Every clause type carries its own question count and its own correct/wrong tally, and answers are marked right or wrong in place rather than totalled at the end.

Practice: clause modules with per-module scores — Practice broken down to the grammar unit, each carrying its own score.

A practice question with inline correct and incorrect marking — Right and wrong marked in place, not tallied at the end.

Phrases, tenses, adverbial clauses and parts of speech — The same pattern repeated across every grammar area.

Shipped

The Courses screen — Courses, for learners who want a structure rather than a streak.

The Translator: text, voice and camera — Translation by text, voice or camera, with a history you can favourite.

Books, Listening Room, Payment and Reading Room — Books and rooms, and the paywall drawn between them rather than assumed.

Profile, reminders, and progress as a chart instead of a single number.

Payment management and email verification — Payment and verification, drawn rather than left to the build.

The LexAI screen set in dark mode — The whole product again, in dark.

Presentation mockups of Courses and the streak — How it was presented.

What I'd change

It won on the strength of the concept, but it was never tested with a learner over more than one session, which is exactly the timescale the motivation argument depends on. I'd treat that claim as unproven until it is.