Home
Scaling Maestro without losing the personal tutor
Maestro is Masterschool’s AI-native learning environment, built around a personal AI tutor for every student. Learning happens primarily through conversation: the tutor explains, asks questions, gives exercises, and adapts as the lesson progresses.
One of the core challenges was making that personal tutor model scalable. Generating every part of every lesson through AI was expensive, so I redesigned the learning experience to combine reusable, pre-generated content with on-demand AI support, reducing repeated generation while preserving the sense of a responsive, personal tutor.
The problem
Every lesson was generated live, turn by turn. That made the experience highly personal, but it also meant generating the same core explanations again and again for students learning the same material.
The challenge was to reduce that recurring AI cost without turning Maestro into a static course.
The solution
We separated the parts of a lesson that could be prepared once from the moments that genuinely require AI. Core explanations, examples, exercises, and learning materials could be generated and reviewed in advance, while Maestro remained available throughout the lesson for questions and personalized support.
The key design challenge was making those prepared materials feel like part of a tutoring session, not like leaving the tutor to enter a separate course experience. Maestro could guide students toward prepared content, hand them an article or exercise, and continue the conversation around it.
The result was a hybrid lesson model: structured where content could be reused, personal where the student needed help.
Make prepared content work inside the conversation
The original chat UI was built around short conversational turns. To support prepared lessons, I adapted the layout so longer explanations could sit naturally inside the conversation without feeling like a separate reading experience.
Guide students through the prepared path
I made prepared follow ups much more prominent, so the intended lesson path was visible and easy to follow without requiring students to keep typing. Freeform input stayed available at all times, so they could still ask their own question or take the conversation in another direction.
Turn prepared content into more ways to learn
Students had already been asking for more ways to learn than chat alone. Once lessons included prepared content, we could turn that into dedicated learning formats like articles, lectures, podcasts, and exercises.
I made those formats easy to access from the conversation, both as suggested next steps and from the plus menu.
Here are a few examples of how that showed up in the product.
Measuring success
The goal was not just to reduce cost. The pilot would only be considered successful if learning quality held up.
- Cost: reduce average AI inference cost per completed lesson by at least 50 percent, including the cost of generating the prepared content.
- Student satisfaction: maintain lesson CSAT at 4.8 out of 5 or higher.
- Progression: keep lesson completion and next lesson progression within 2 percentage points of the existing experience.
- Learning quality: keep mastery checkpoint performance within 2 percentage points of the control.
