Gabriel Kasmi

Gabriel Kasmi

Applied AI Scientist

Welcome

I am an Applied AI Scientist working at the intersection of deep learning, remote sensing, and energy systems. My PhD and postdoctoral research at Mines Paris - PSL focused on the reliability and transparency of deep learning models deployed in industrial settings, with applications to the remote sensing of rooftop photovoltaic (PV) systems.

I created and maintain DeepPVMapper, an open-source system that has mapped ~500,000 rooftop PV installations across France and enabled the audit of RTE's grid connection data. My methodological contributions include work on explainable AI published at ICML and computer vision applied to Earth observation.

To prove this research could hold up as a real product, I co-founded Solar Dashboard, taking DeepPVMapper from a research pipeline to a tool used by industry players — concrete evidence I can carry work from paper to production.

Bienvenue

Je suis Applied AI Scientist, à l'interface du deep learning, de la télédétection et des systèmes énergétiques. Ma thèse et mon postdoc à Mines Paris - PSL portaient sur la fiabilité et la transparence des modèles de deep learning déployés en environnement industriel, avec pour cas d'application la télédétection des installations photovoltaïques (PV) sur toiture.

J'ai créé et je maintiens DeepPVMapper, un système open source qui a cartographié ~500 000 installations PV en France et a permis l'audit des données de raccordement de RTE. Mes contributions méthodologiques incluent des travaux sur l'IA explicable publiés à ICML et en computer vision appliquée à l'observation de la Terre.

Pour prouver que cette recherche pouvait tenir la route en tant que produit, j'ai co-fondé Solar Dashboard, qui transforme DeepPVMapper en un outil utilisé par des acteurs de l'industrie — la preuve concrète que je peux porter un travail de la publication à la production.

News

  • March 2026 A new preprint on model efficiency for rooftop PV detection is available here. Using data from Madagascar, this work challenges the "bigger is better" assumption in computer vision: smaller, high-resolution configurations consistently outperform larger models in data-scarce Earth observation settings.
Highlights:
  • July 2025: I presented our feature attribution method WAM at ICML 2025. Access the paper in the research section.
  • April 2025: Our latest paper has been published in Environmental Data Science. The paper is accessible here.
Past events: