François-Guillaume Fernandez · Paris
01 · Interpretability
TorchCAM scores every cell of the network's last grid to show what drove the prediction: here, the head and chest. It started as a project to better understand and implement research papers on computer vision explainability.
Photo: Woopets.
02 · Food recognition
Foodvisor turns a meal photo into nutrition tracking: each food is found region by region, then becomes a portion with its calories.
03 · Document understanding
Three dates on the page, one due. Reading an invoice means finding the text, decoding every character and knowing which field it belongs to.
04 · Small-object detection
The earliest smoke is the smallest: here, 23 × 29 pixels, less than one 32-pixel cell. Spotting it is half the job; firefighters also need its GPS position, from the bearings of two cameras.
Frames and labels: Pyro-SDIS, Pyronear's open dataset (camera cabanelle-125, 24 Feb 2024).
05 · Video & intent
One frame shows a person; a few seconds show intent. Rimward detects someone entering a protected site and says where: which site, which camera, which stretch of fence.
Intrusion detection for solar plants.
It started with a question about copper theft.
A site we were protecting against wildfire with Pyronear asked whether we could help with copper theft. That request became Rimward. I founded it and built the product end to end.
Initiator & co-founder · 2019—
Smoke, 30+ km away
4–5 cameras per station · detection on site
Wildfire detection people can afford to deploy.
I pitched the idea to a volunteer association and built the first vision models and backend API. To find training data from the field, we collected imagery from US national-park cameras rather than Google Images. We chose affordable cameras and straightforward deployment over satellite imagery.
The wider team now develops the vision algorithms. The network has reached France, Chile, Spain and beyond; in summer 2026, a camera detected a wildfire more than 30 km away.
Research papers. A small codebase. An API I wanted to use.
I built TorchCAM on my own as a researcher to look for visual bias in the models I trained. The existing tooling wasn't the interface I wanted, so I implemented the research papers behind class activation maps in a small library with a clean PyTorch API.
It has since found its way into research, including COVID-19 chest X-ray studies. The published demo shows a real class activation map.
On science, technology and creativity.
Frontend finally clicked: a Python-first mental model
2025-10-25Detect forest wildfires using AI ↗
Datagen podcast
Early detection of wildfires ↗
DataXDay 2021
Beyond the projects
Books, films, tools, and the things I keep coming back to.
A look at my desk →