Rebuilding onboarding around trust
Summary
As the sole product designer at Heylama, I reworked onboarding end-to-end, from the first screen to the learning plan. We were asking users to trust the product before showing them why they should. Rebuilding around that doubled conversion.
Team
Timeline
Mar–Jul 2025
Impact overview
Maybe the problem wasn't the screens.
Conversion rates had been sliding, and the team had already tried a few onboarding changes. None of them held.
Before changing the UI again, I wanted to understand what the existing flow was asking users to do and believe. That meant looking beyond individual screens and at the experience end-to-end.
Diagnosis
We were asking for commitment before showing enough value.
Mapping the onboarding, reviewing LogRocket sessions, and auditing each decision before users reached the product exposed a broader pattern.
The flow kept asking for effort, information, and trust before users had enough evidence that the product could help them. Five friction areas kept showing up.
Benchmarking
Was this a Heylama problem or the whole industry's?
Comparing Heylama with other language-learning products surfaced a consistent pattern: stronger onboarding experiences established value and confidence before asking users for more effort.
The point was not to copy a specific flow. It strengthened the case that sequencing, not just screen design, was part of the problem.
A system built to earn trust first.
The first direction treated onboarding as one connected experience rather than a sequence of setup screens.
Value came earlier. The tutor became something users could interact with. Personalization had to produce a visible response. The AI had to demonstrate what it could do. Progress needed to feel real before signup.
Each part addressed a different piece of the same trust problem.
Constraint
The full idea never made it into the product.
Only part of the direction shipped in the first sprint. The value proposition moved earlier, but the conversation, visible personalization, proof, and learning plan did not come with it.
When a React Native upgrade broke core parts of the app, fixing the product took priority. What remained was a much smaller version of the concept. Conversion fell to 4.3%, and the team rolled back to the last stable release.
The direction was still unresolved. We had changed parts of the experience without getting the chance to build the system behind them.
Testing the version we could actually build.
After the rollback, rebuilding the full concept was still beyond the team's capacity. The existing onboarding became the base for a more focused revision built around the same principles.
That revised prototype went into moderated user sessions. The goal was to understand when the product started feeling useful, whether personalization felt credible, and whether users understood where the onboarding was taking them.
Insights
Three signals changed what we focused on.
Testing made the trust problem more specific. Personalization felt fake when user input disappeared into the flow. The “how it works” screens confused people because they looked interactive without actually being interactive. Users also responded better when onboarding gave them something concrete to aim for.
AI helped cluster recurring themes across the transcripts, with every finding checked against the original recordings before it informed a design decision.
One extra hunch went into testing too: replacing the human tutor with a consistent llama character. Participants described the llama version as less judgmental and more relaxed. That small experiment would later grow well beyond onboarding.
Turning research into the highest-impact
changes.
With the direction clearer but engineering capacity still limited, the rebuild focused on the moments where the existing flow could carry the research further.
Personalization
Goals became skills.
The old question mixed broad motivations like work and travel with specific outcomes like preparing for a presentation. The result was a long list where users often chose the closest match rather than one that felt true.
Replacing those options with concrete skills reduced decision fatigue and gave the product input it could act on directly. Each choice mapped to the learning experience, making the reason for asking much clearer.
Some motivational context was lost, but the answers became easier to choose and more useful to the product.
Sequencing
Small changes fixed important moments in the flow.
Adding an A0 beginner level gave complete beginners somewhere accurate to start. Before that, A1 was the lowest option, which could make someone struggle immediately and feel behind from the first lesson.
Social proof moved earlier, before users had invested much effort. The feature tour moved later, once users had enough context to understand why those features mattered to them.
Each change was small. Together, they made the sequence easier to understand and easier to trust.
Karl
A research finding became the brand.
Before this project, llama illustrations appeared across Heylama with no consistent look, role, or voice. Anka, the human tutor, also appeared throughout the experience, which made the product identity feel disconnected.
Testing showed that the consistent llama character made the interaction feel less judgmental and more relaxed.
That finding grew beyond onboarding. Karl became Heylama's official mascot, with a defined appearance, personality, tone, clothing, and rules for how he should behave in the product. He later became the app icon and appeared across product, marketing, and social.
Learning plan
Personalization needed a visible payoff.
The old flow collected goals, level, and preferences, but users had little proof that those answers affected anything.
The learning plan turned that input into something visible. Users could see their starting point, where they were heading, and how the product planned to help them get there.
Every question creates an expectation that the answer will matter. The plan showed users where their effort went.
The final version doubled conversion.
The research pointed further than we could ship.
Some of the strongest ideas stayed in prototype:
- Live level check
Replaced self-reported level with demonstrated ability, giving the product a more reliable starting point. - AI tutor conversation before signup
Let users experience personalization before asking them to trust it. - Product experience before account creation
Moved the first moment of value ahead of commitment.
Those ideas needed more engineering capacity than we had. The shipped version kept the existing structure and concentrated the research into the moments we could change with confidence.
The project also changed how I read funnel problems. Analytics can show where someone leaves, but that point is often only the consequence of something that happened earlier.
Drop-off tells you where people leave. It does not tell you when they stopped believing.
