The work
Physicist by training, AI builder by trade. Nine years building the AI behind Legit.Health’s CE-marked medical device. Research on learning from very few images, from plant disease to histology. Different fields, the same method.
Legit.Health, 2017–2026.
I was co-founder, Chief AI Officer, and Technical Director & PRRC.
Context
Dermatologists are few and waiting lists are long. The severity scores that guide treatment are filled in by hand, and specialists often disagree on them.
Constraint
Data. Good labelled skin images are scarce, and experts disagree. The labels had to come from a consensus of several specialists, or the model would learn their bias.
Regulation. The product is a medical device. Under the EU MDR, every model needs evidence, documentation and a controlled release, inside an ISO 13485 quality system.
What I led
I led the vision engine that grades skin disease from a phone photo. It recognises more than 300 conditions and automates the severity scores dermatologists use, such as PASI for psoriasis and SCORAD for atopic dermatitis.
I chose the metric for each use case first, then worked on the dataset, then on the models.
How it was validated
Multiple clinical validation studies. CE marking: first as class I under the former Medical Device Directive, and since September 2026 as class IIb under the EU MDR, certified by BSI, notified body 2797 (certificate), with an ISO 13485 quality management system.
I was Technical Director & PRRC (the person responsible for regulatory compliance). I led the AI process and took part in the clinical work.
Outcome
- 2,500+
- clinicians
- 300+
- skin conditions
- $7M
- raised by Legit.Health
“…improving diagnostic accuracy, cutting consultation times in half and reducing waiting lists.”
Forbes, 30 Under 30 Europe 2025, on Legit.Health What this means for you
I have done the whole path once: data, model, evaluation, clinical validation, CE marking.
On your project that means the metric is agreed before the model, the evaluation is one you can show an auditor, and the documentation is written as the work happens.
Research.
Few-shot learning: models that learn from tens of examples, not millions.
- 847
- citations
- 29
- papers
- 10
- h-index
My most-cited paper (435 citations) applies few-shot learning to plant disease from field photos. It was research, not client work, and it shows the method works outside medicine.
Five photos per class, in your browser.
Most companies don’t have a million labelled images. For many tasks, you don’t need them.
Runs in your browser. Your photos never leave it.
Book a 30-minute call to check fit.
The call is free. I reply within two working days.