Version 5
Lessons felt more static, with limited interaction states, weaker recovery moments and less guidance when learners became stuck.
01 Case Study
Kaizen Languages.
Designing an AI-powered language learning experience around confidence, motivation and retention.
As founding product designer, I led the end-to-end design of a native iOS and Android language learning app, shaping the lesson experience, onboarding, writing interactions, gamification, experimentation and conversion journeys. The challenge was not only helping people learn a language, but helping them feel confident enough to keep going.
- Role
- Founding Product Designer
- Timeline
- 3 years
- Platform
- Native iOS and Android
- Team
- Founders, developers, language specialists, AI stakeholders
- Focus
- Mobile product design / Behavioural UX / AI conversation / Retention / Gamification / Subscription conversion
+25%
Lesson completions after Suggestion
3.8%
Conversion at Version 7
+35%
Retention uplift through gamification
62% NPS
Writing feature score
02 Overview
Designing an AI learning product people would return to.
Kaizen was built around a different way of learning languages: AI-powered conversation instead of traditional classroom or video-based learning. That created a complex product challenge. Users needed to understand how the product worked, feel comfortable speaking to AI, recover from mistakes and stay motivated long enough to build a habit.
As founding product designer, I shaped the mobile experience from early discovery through optimisation, working across lesson design, onboarding, conversion, analytics, gamification and writing-based learning systems.
My role
I owned the product design from discovery through optimisation — across UX, UI, research, analytics, workshops, guidelines, product foundations and hands-on delivery. Because Kaizen was an early-stage product, the role required both strategic thinking and execution.
- Defining the core lesson experience
- Building testable prototypes
- Creating product flows and UI systems
- Running guerrilla and remote user testing
- Using Firebase funnels and Hotjar to analyse behaviour
- Shaping A/B tests and CRO decisions
- Producing assets and documentation for developers
- Collaborating with language specialists on lesson structure and writing systems
- Working with developers on speech, writing input and interaction constraints
03 The challenge
Creating trust and momentum in a new way of learning.
The core challenge was not simply helping users complete lessons. It was helping them build confidence. Speaking to AI felt unfamiliar and high pressure. When users made pronunciation mistakes, repeated correction often created frustration and drop-off. The product needed to feel educational, responsive and encouraging without removing the challenge of learning.
How might we design a language learning experience that feels adaptive, rewarding and behaviourally sustainable over time?
- Speaking to AI felt unfamiliar and high pressure
- Users dropped out when pronunciation errors repeated
- Early lesson experiences were not interactive enough
- The product needed to balance learning outcomes with subscription growth
- As an early-stage product, we needed to learn quickly through research, testing and analytics
05 Discovery and research
Understanding how people learn, where they struggle and what keeps them going.
To design a credible learning product, I needed to understand both the educational and behavioural side of language learning. The research focused on how learners build confidence, where they lose momentum and what type of support helps them continue after making mistakes.
To avoid testing with generic app users, I recruited people actively learning Japanese through language schools, meetups, comic stores, WeWork, flyers outside independent classes and remote platforms such as UserTesting and TryMyUI. This helped validate the product with learners who were already experiencing the motivation, confidence and pronunciation challenges the app needed to solve.
Research inputs
- Competitive review of language and chat-based products
- In-person guerrilla research with learners
- Remote testing using platforms such as UserTesting
- Visits to language schools, meet-ups and learning communities
- Study of spaced repetition and language memory techniques
- Study of beginner Japanese learning methods, including Genki
- Collaboration with language specialists and AI stakeholders
- Firebase funnels and Hotjar for behavioural and drop-off analysis
- A/B testing across onboarding, lessons and subscription screens
What I was trying to learn
- What makes a lesson feel useful rather than intimidating
- Where users lose confidence
- How much guidance is enough without reducing challenge
- What drives repeat use and paid conversion
- Whether beginners needed more foundational support before advanced content
04 Key user insight
Learners did not just need content. They needed confidence.
The biggest friction was not simply “learning a language”. It was the emotional barrier inside the lesson. Users were often unsure whether they were pronouncing things correctly. When correction repeated without enough support, the experience started to feel punitive rather than helpful.
Confidence mattered as much as correctness
Repeated correction created frustration
Users needed help recovering when they got stuck
The lesson flow needed to support progression, not just identify errors
Feedback needed to feel encouraging, contextual and actionable
Beginners needed a stronger foundation before being pushed into more advanced content
Learners were not dropping out because they lacked content. They were dropping out because the experience did not always help them recover when confidence broke down.
06 Core lesson experience
Making AI chat feel teachable, responsive and engaging.
One of the core product problems was how to translate a chat interface into an effective lesson environment. The messaging experience had to do more than look like chat. It had to teach. I explored how users would distinguish AI responses from their own, replay audio, view translations, understand errors and hear slowed-down pronunciation.
- Differentiated chat bubbles and message states
- Typing and audio feedback to make AI feel responsive
- Translation controls
- Character-set support such as romaji and hiragana
- Instructional cards and correction states
- Prototypes and documentation for developer handoff
Where the lesson experience broke down
Through guerrilla testing and recorded sessions, I found a recurring issue: users often believed they were saying the right word, but still received errors. After repeated failed attempts, frustration increased and lesson drop-off rose sharply. Correction alone was not enough support.
Users needed help recovering from mistakes, not just being told they were wrong.
07 Design decision
Helping users recover when pronunciation failed
To reduce frustration, I introduced Suggestion, a support card that broke pronunciation into simpler syllables and allowed users to hear the word more slowly. The goal was to preserve challenge while giving users a way through moments of failure. The design was triggered after repeated errors, directly inside the lesson flow.
- Reduced dead ends in the lesson
- Gave users a clearer path to success
- Supported confidence without removing learning effort
- Kept help contextual rather than hidden elsewhere
Suggestion gave learners a clearer recovery path when pronunciation failed, without removing the learning challenge.
08 Iteration
From static lessons to interactive learning loops
Version 6
Version 6 introduced AI typing states, clearer options, drawer cards, grammar prompts and more supportive recovery moments, making the experience feel more responsive and easier to continue.
Conversion rose from 1.6% in Version 5 to 2.7% in Version 6, a roughly 70% uplift, with Version 7 later reaching 3.8%.
09 Product strategy
Why I prioritised beginner confidence over more advanced content.
A key strategic moment in the project was challenging the initial direction from the CEO and CTO. The business wanted to add more intermediate and expert-level lessons to increase revenue and reduce churn. But research suggested that our strongest opportunity was not simply adding more content.
Most of our audience was still at the novice or beginner stage. Before asking users to progress into harder lessons, they needed a stronger foundation in the writing system itself. I proposed shifting focus towards Hiragana and Katakana, creating a feature that helped beginners learn character recognition, stroke order and recall through guided interaction.
This moved the product strategy from “add more lessons” to “strengthen the learning foundation”.
Design decision: Prioritise foundational confidence before advanced progression.
Challenged the assumption that more advanced content would reduce churn
Used research to identify beginners as the higher-impact audience
Proposed a foundational writing feature focused on Hiragana and Katakana
Positioned Kanji as a later progression once the learning model was validated
Connected beginner confidence to long-term retention and subscription value
Design judgement
Designing for moments when AI breaks down.
Conversational learning depends on confidence, but speech recognition and repeated corrections can quickly create frustration. I focused on recovery moments that helped learners continue without feeling punished by the technology.
Suggestion cards after repeated errors
When learners struggled to pronounce unfamiliar words, I introduced a suggestion card that broke difficult words into syllables and let users hear slowed audio before trying again.
A skip path when recognition failed
Analytics showed that some learners were leaving lessons when microphone recognition failed repeatedly. I introduced a skip option after a minimum number of corrections so users could continue without abandoning the lesson.
10 Writing system
Translating a physical learning method into a digital experience.
To design the writing feature, I first needed to understand how Japanese characters are traditionally learned. I studied Genki and beginner learning methods to understand how Hiragana and Katakana are introduced through character groups, repetition, stroke order and recall.
The goal was not simply to show users a character. It was to help them understand how each character is formed, practise the movement and then recall it later without support. I broke the learning journey into character sets, starting with the first Hiragana group: A, I, U, E, O.
Learning model
- Introduce the character visually
- Show how the character is written through animated stroke order
- Let the user practise with a dashed guide
- Reduce guidance by removing the dashed line
- Increase difficulty by removing start and end markers
- Test recall through a short quiz
- Reward completion to reinforce progress
Each stroke used a green start point and a red end point to guide the user. The dashed line showed the correct stroke path at the easiest level. As the user progressed, the help was gradually removed so they could build confidence and memory.
Level 1
Dashed line, green start point and red end point are visible.
Level 2
Dashed line is removed. Only the green and red points remain.
Level 3
All guidance is removed and the user has to recall the character independently.
At the end of each character set, users completed a short quiz combining characters they had learned. For example, after learning A, I, U, E, O, users were shown AO, meaning blue in Japanese — helping users feel that the writing exercise had practical value beyond isolated character practice.
I worked closely with language specialists to define the character sets, pronunciation and quiz logic. I also collaborated with developers to understand the technical constraints of stroke recognition, including the acceptable drawing area, stroke size and preventing users from completing the interaction by drawing outside the required path.
62% NPS in testing, 12% above average. In testing, users were highly engaged with the writing mechanic and were able to recall the majority of characters during the final quiz.
The writing feature shifted the experience from passive learning into active practice, giving beginners a clearer sense of progress and achievement.
11 Gamification
Designing motivation through behavioural psychology.
To improve retention, I explored how successful learning products sustain engagement over time. Through competitor analysis, particularly studying Duolingo, it became clear that gamification is not just a feature. It is a behavioural system.
- Reward loops reinforced repeat actions
- Progress visibility created achievement and momentum
- Loss aversion encouraged consistency through streaks
- Goal-setting mechanics increased commitment over time
I introduced a gamified system designed to support daily learning behaviour, including streak mechanics, reward structures, progress markers and future-facing ideas such as streak freezes, repairs and leaderboards.
Visual system
- Lightning bolts represented energy, momentum and streak continuity
- Stars acted as immediate rewards for lesson completion
- Trophies signalled achievement and future leaderboard potential
Writing exercise connection
- Difficulty increased as guidance was removed
- Quizzes reinforced recall at the end of each character set
- Rewards helped users feel progress after completing a learning loop
- The mechanic supported both motivation and memory
12 Conversion
Conversion and subscriptionexperiments.
Balancing user value with subscription conversion
I used analytics and A/B testing to understand how lesson access influenced subscription behaviour. One experiment tested whether offering 2 or 3 free lessons per day led to stronger premium conversion. The data showed that limiting access to 2 free lessons increased subscription conversion from 1.15% to 1.3%, resulting in a 13% increase in premium customers.
This was not about adding friction for the sake of conversion. It was about finding the right balance between giving users enough value to understand the product and creating a clear reason to subscribe.
+13% increase in premium customers.
Small hierarchy changes created measurable subscription uplift
Another experiment focused on the subscription screen, specifically the placement of the Most Popular label. By testing which plan was framed as the recommended option, the team found that moving the label increased 6-month subscriptions by more than 20%. It showed that not every gain requires a major redesign. Sometimes clearer framing changes decision-making.
More than 20% increase in 6-month subscriptions.
Reducing sign-up friction through social login
I also pushed for lower-friction sign-up patterns, including social sign-in. This was initially debated with stakeholders, but the rationale was clear: fewer fields and fewer decisions reduce friction at the point of entry.
- Sign-ups increased by up to 40%
- Subscription rates increased by 20%
- Social links on app performed 8.5% better for conversion than web
Small interaction changes mattered because the product depended on momentum. Every moment of friction risked breaking the learning habit.
Proof
What changed the product direction.
Research, moderated testing and product analytics challenged the assumption that more advanced content alone would improve retention. The strongest gains came from recovery design, beginner foundations and small, measurable experiments.
Completion improved through character control
Giving learners the option to switch between Romaji, Hiragana and Kanji gave them more control over difficulty and increased lesson completion by 25%.
Writing validated active learning
The writing feature was shaped around stroke order, repetition and recall, then validated with a 62% NPS, 12% above average.
Small pricing changes improved conversion
A/B testing showed that moving the ‘most popular’ label to the 6 month plan increased 6 month subscribers by more than 20%.
The product direction shifted from adding more content to designing the confidence loops that helped learners keep going.
13 Impact
Impact at a glance.
The writing feature became one of the strongest indicators that beginner-focused learning could create deeper engagement. It showed that users valued active practice, guided progression and the feeling of genuine learning achievement.
Alongside moderated testing, I used Firebase funnels, Hotjar and A/B testing to understand where learners dropped off, which lesson changes improved completion, and how pricing presentation affected subscription behaviour.
14 Decisions
The decisions that shaped the product.
15 Reflection
What this project taught me.
Kaizen showed me that successful learning products are not simply educational systems. They are behavioural systems.
Retention was driven less by content volume and more by confidence, emotional reinforcement, progress visibility and habit sustainability.
The project shifted the design challenge from “How do we teach language?” to “How do we help people continue learning?”
Designing for learning behaviour required balancing
- Structure with flexibility
- Guidance with experimentation
- Progression with emotional motivation
- Learning outcomes with commercial sustainability
- Business goals with the needs of the core learner audience
This project shifted the product from simply adding more content to designing the confidence loops that helped learners keep going. The biggest improvement came from combining research, analytics and small product experiments to reduce friction, support recovery and make progress feel more achievable.
16 Retention
Turning first-time use into repeat behaviour.
Once the core lesson experience improved, the next challenge was retention. I looked at how learning products create momentum and studied gamified patterns used by competitors. This led to the introduction of streaks and reward systems that encouraged daily return behaviour.
- Retention increased by over 35% after introducing streaks
- Kaizen achieved 40% Day 1 retention versus an education-tech average of 14.5%
- Kaizen achieved 22% Week 1 retention versus an average of 4.6%
