Helping QA learners learn more confidently with an AI tutor built for trust
MY ROLE ON THE PROJECT
Senior Product Designer, solo on this project. I led the 0 to 1 design of Ela, QA's AI learning assistant, working closely with Product, Engineering and the internal AI team.
My objective was to turn an AI assistant from a scattered coding hint tool into something learners would actually rely on across their whole course, without losing trust in the process.
Timeline: 2 months to MVP. Launched January 2025.
Leadership wanted to bring AI to the whole platform as quickly as possible. I focused on where it could genuinely help learners.
QA already had a small AI assistant living inside coding labs, but it only helped with one narrow task and felt disconnected from everything else learners did on the platform.
AI was quickly becoming something learners expected everywhere, not just in one corner of a course, and leadership wanted it across the whole platform.
The starting point was an early prototype with real UX and technical limits.
Our team had little experience developing AI products, and they gave us under two months to ship something meaningful so we were forced to make fast, pragmatic choices.
The real question was: What would a genuinely useful assistant look like for someone trying to learn
Especially given that AI used in education is treated as a higher risk category under frameworks like the EU AI Act: the people affected are often vulnerable, and a careless or wrong answer costs more than in a typical app.
What actually helps a learner who is stuck?
Before designing anything, I looked at how tools like Duolingo and Khan Academy worked, alongside academic research from Harvard and Vilnius University on AI assisted education. And some patterns started to emerge.
I noticed that the most effective learning assistants included:
I paired that with mapping the learner journey directly on QA's platform, and found four recurring moments where people struggled:
Coming back to a course after a break, getting stuck on a difficult concept, preparing for exams, and reinforcing what they had just learned. I wanted Ela to appear at the moments learners needed support, with help that fit what they were doing.
I chose guided actions over open chat to make the experience clearer for learners and more predictable for the team.
At the time, open chat was becoming the standard way to build AI assistants. There was real enthusiasm internally for giving learners a free-text chat like the ones they were seeing elsewhere, and it looked like the fastest option to build. But my research and Engineering’s concerns pointed towards a more guided experience. I designed Ela around Smart Actions: a small set of predefined ways to get help, so learners knew what they could ask and what to expect. This also made Ela’s behaviour easier for the team to explain and support beyond a demo.
So the first version of Ela was born
To turn that decision into something learners could actually use, I designed a small set of structured actions built around what I had already found they needed most: a way to recap, a way to check their own understanding, and a way to get something explained.
Summarize
Helps learners quickly recap lessons and reinforce retention.
Test
Generates quizzes that enable self-assessment and strengthen knowledge.
Explain
Breaks down complex concepts using clear, step-by-step guidance or analogies. These features transformed passive learning into an interactive and supportive experience.
Designed to blend into the learning experience
For the assistant to feel useful rather than intrusive, I designed it to function as a companion throughout the existing learning experience:
Embedded within lessons and articles
Accessible during key moments in the learner journey
Connected to QA’s existing design system
Enabled through contextual interactions such as text highlighting for instant explanations
Reusing the old prototype made more sense than starting over
The second real choice was how to treat the inherited proof of concept. With less than two months to reach MVP, rebuilding from scratch would have left little time to ship. I reviewed the prototype and kept the parts that gave us a sound starting point.
That choice let the team move fast while still building on solid ground rather than a shortcut, and it meant the components I shipped were built to scale into future AI features on the platform, not just to patch the current one.
The numbers behind a gradual launch
Ela rolled out gradually between January and July 2025, with very little promotion behind it.
The numbers below reflect usage data updated as of September 2026:
It has since grown to over 150 client accounts,
with roughly 700 to 1,000 learners a month actually sending it messages, not just opening it.
Conversations run about 4 messages deep on average, a small but real sign that learners are using it for more than a quick glance.
Around 29% of first time users return to Ela within their first month, and it holds a 4.5 out of 5 satisfaction score with an average session of 14 minutes.
Beyond the numbers, the decision to hold the line on guided interactions rather than open chat turned out to be the right call. Learners engaged consistently with the structured actions, and the constrained design gave both the business and the learners something they could trust from day one, which mattered more at this stage than raw conversational flexibility would have.
Watching how learners actually used it
Once Ela was live, I looked at natural language interactions and ran live sessions with apprenticeship learners to understand real usage rather than relying on assumptions. A few patterns stood out:
Learners under time pressure leaned heavily on summarization
Many turned to the assistant specifically to prepare for exams
A good number were looking for reassurance and direction as much as for information
Learners seemed to want Ela to guide them proactively, especially when they were preparing for exams or needed reassurance.
Looking back
Ela started as a constrained MVP built under real organizational and technical pressure. By choosing guided interactions over open chat, and treating that choice partly as a responsible AI decision rather than only a design one, I helped QA ship an assistant that learners trusted quickly and that the business could stand behind.
Since launch, Ela has expanded across nearly the whole platform, with one deliberate exception: it was kept out of assessment content, since letting learners lean on it there would have undermined the point of the assessment itself.
I also added source citations to its answers, so learners could see where an explanation came from rather than just taking the assistant's word for it, which mattered more as Ela's footprint grew.
More than the immediate results, the project set a pattern for how QA approaches AI going forward: as structured, accountable support built into the learning experience, not a novelty layered on top of it.
