An AI Ballet Coach Built to Let You Fail Comfortably
TL;DR
Ballet teachers teach the way they were taught — when they started at the age of 5 and aim to perform professionally. I'm building an AI coach for the much larger group it ignores: adults who start late, want to genuinely improve, and need something the industry has never designed for. Not a better teacher. A safer place to fail.
My Role
Founder & Designer — Ballet has been my passion as both learner and teacher. AI makes it possible to bring the patience, repetition, and psychological safety of an exceptional one-to-one teacher to every adult learner.
01 The Gap
Most ballet teachers began around the age of 5, absorbed technique through repetition before they can question it, and were trained to demonstrate — not diagnose. That's a reasonable model for professional-track students. It's a poor model for adult recreational learners, who have different bodies, different timelines, and a decade of compensations already built in from elsewhere.
The industry serves one student type and treats everyone else as a lower priority.
02 The Real Problem
The technical gap matters less than the psychological one underneath it.
Adult learners carry self-consciousness a 5-year-old doesn't. In a group class, a correction is public. Getting something wrong in front of people who started decades earlier feels expensive — so adults under-try. They hold back from full commitment to the movement precisely because failing fully feels humiliating.
But full attempts — including the failed ones — are how technique is actually learned. Meanwhile, a teacher in a room of 20+ can give any 1 student perhaps 90 seconds of real attention per class. The two halves of the learning loop — psychological safety and repetition volume — are structurally missing.
03 The Design Approach
The MVP is scoped to one movement: the pirouette. It's the most mythologised step in ballet, the one adults fear most, and the one with the clearest mechanical basis.
I'm treating this as two separate design problems, not one product:
The correction engine
A student uploads video of their turn. The system identifies the single most significant deviation from a physically grounded reference, explained in plain terms rather than ballet jargon. One diagnosis, one explanation, one at a time.
The pedagogical behaviour layer
This is the harder and more important problem. The system needs an explicit point of view on tone: patient rather than clinical, encouraging failure rather than merely tolerating it, treating repetition as the point rather than a symptom of not getting it. Holding both precision and warmth simultaneously is a great design challenge.
04 Assumptions to Consider
There are 2 underlying assumptions that are expensive to be wrong about after infrastructure is built — but cheap to test.
1: Would a "single most significant deviation" framework actually work on real, messy adult bodies?
2: Would an encouraging tone change whether someone attempts again, rather than just feeling nicer in the moment? People have different learning styles, different relationships with failure, different responses to correction. MBTI, attachment theory, and learning style research all suggest the optimal feedback model may need to adapt to the individual rather than apply uniformly — a fun research question!
For now, I need a phone camera, real people willing to fail at a pirouette, and enough structured observation to find out whether the core framework actually works. The ML and the app come much later, once the concept is proven with almost no code at all (hopefully!)
To Be Continued