Data scientist

I teach machines to notice what's off.

🔍 Théo Taburet · data science · Antibes

I’m Théo Taburet, a data scientist at Ezako in Antibes, France. I build models that find the one thing worth looking at: a sound in hours of underwater recordings, a sensor whose signal stops looking normal, a retouched patch in a photograph.

At Ezako I build detectors of the YOLO and RF-DETR family that annotate underwater sound, and anomaly detection for MEMS sensors: 1D CNNs, quantised until they fit on an STM32 microcontroller.

Before that, six years of images. A PhD in applied mathematics at École Centrale de Lille on steganography: hiding a message in a photograph so that it passes for the camera’s own noise. Then a postdoc at Université de La Rochelle on the other side of the coin: spotting retouched receipts and double-compressed JPEGs. The thesis’s main result now has an interactive explainer.

What I work on

Machine learning: deep learning, CNNs, object detection (YOLO, RF-DETR), anomaly detection, segmentation
Embedded ML: quantised 1D CNNs on STM32 microcontrollers
Signals and images: underwater sound, MEMS sensors, image processing, image forensics
Tooling: Python, PyTorch, OpenCV, FastAPI, Docker, CI/CD

What I’m building

diapason is your own radio, broadcast from a phone. One person picks the music, everyone around scans a QR code and plays it out loud on their own portable speaker, in sync, with nothing to install. A Rust server, a WebAssembly client, and a lot of care about clocks.

The projects page has it in full, with the rest: a photograph printed in four trembling inks, and a burst of photos turned into a looping video, both in the browser.

Away from the screen

I ride bikes, usually for longer than is reasonable. I co-founded Pignon des Charentes, which runs riddle-based bike treasure hunts around La Rochelle. I also built Ravitools to solve a problem I kept having on long routes: knowing where the next water source is when you have no signal.

Both of those and diapason are the same itch, really: making a group of people share something at the same time, in the same place.

And I take photographs.

Good atSpotting the one point on the curve that has no business being there.

Hopeless atStopping before the 200th kilometre.