Product
AI iPad App & Web Resources
Children complete assessments and training in the app with additional resources to print extending the experience beyond the screen.
Case study · Education App
Dynamilis
Transforming an EPFL research prototype into an iPad app with 300K+ downloads. Top 5% retention and conversion in EdTech.
Co-Founder
& Head of Product
2020
– 2026
Overview
From EPFL research to App Store success
Team & context
EPFL Spin-off
Started as a research project at EPFL, one of the world's top-25 science and technology universities, to become a startup with a cross-functional team of ~10 and school partnerships across five countries.
My Role
Co-Founder & COO, Head of Product
Owned roadmap, experimentation, UX, acquisition (paid & organic), and built the product organization from scratch.
The opportunity
Bringing handwriting assessment from the lab to the living room
The friction
A hidden crisis
Handwriting difficulties affect 10–30% of children but are often detected too late. Teachers lack time for individual assessment, and parents struggle to track progression objectively.
The starting point
A brilliant research prototype
The core technology existed at EPFL, but with no users and a small dataset (=low reliability). The interface was strictly built for researchers, too complex for families or schools to use independently.
The challenge
Crossing into the real world
We had to answer fundamental questions: What's the simple value proposition for parents? How does it fit in busy classrooms? How do we build a massive robust dataset across European writing systems?
How it works
A simple loop for children, parents and teachers
Users capture handwriting via iPad using a pencil, the AI assesses it in real-time, and targeted minigames are prescribed.
Execution
Hypothesis-driven decisions, shipped fast
Product Strategy
Define the core loop
Instead of exposing every research capability, I defined a single loop: capture handwriting → analyze → guide targeted training. This became the north star for design, engineering and GTM.
Discovery & delivery
Own the backlog
Bridged feedback from parents, therapists, and behavioral data into clear cycles for a cross-functional squad of 10. Every sprint shipped meaningful, measured improvements.
Data Strategy
25K-child dataset
Led partnerships with 50+ schools to build a massive, real-world European dataset. This deeply improved model accuracy while proving viability inside classrooms.
Go-to-market
Marketing & Sales
Created App Store assets, ran Meta/Apple Search Ads campaigns, and closed deals with Swiss schools. Built scalable acquisition channels with minimal budget.
Highlight
App Store conversion: 1.5% → 6.5%
To address a weak impression-to-install rate, I ran structured A/B tests over six months. With the same underlying product, changing screenshots and copy took conversion from 1.5% to 6.5% (a 4.3× increase).
Before – (1.5% conversion)
After – New narrative & visuals (6.5% conversion)
Result
The numbers that prove it works
300K+
Downloads
Monthly retention and download-to-paid conversion ranked in the top 5% among 75,000+ apps tracked by RevenueCat, in one of the toughest app categories for monetization.
+21%
Learning speed
In partner-school cohorts, children using the app regularly improved their handwriting speeds significantly faster than control groups.
Top 5%
Growth efficiency
App Store funnel efficiency ranked in the top 5% among 405,000+ education apps, achieved on a low budget through organic trust and structural ASO optimizations.
25K+
Strategic moat
Created one of the largest structured tablet-handwriting datasets in the world, driving continuous ML research and raising product barriers to entry.
Retrospective
What I learned
Research ≠ product requires brutal simplification
Early versions tried to show everything the system could detect. Users were confused. The breakthrough came when we focused on one clear promise and hid complexity behind progressive disclosure. Sometimes the hardest product decision is what to leave out.
B2B and B2C need different truths
Parents wanted reassurance and fast results. Schools needed proof it wouldn't disrupt day-to-day reality in the classroom. Same product, completely different positioning. In hindsight, I would focus on nailing one motion perfectly before trying to grow B2B and B2C in parallel.
Data partnerships are product work
Building the 25K-child dataset wasn't just about model accuracy, it shaped our understanding of real constraints, built credibility with schools, and forced us to design for messy, real-world conditions instead of lab perfection. The product and the research evolved together.