I build machine learning and computer vision systems, often in places where the data is messy, labels are non-existent, and the obvious solution is wrong. My background is in geospatial AI, scientific computing, and reproducible ML workflows. These days I am trying to push that work further toward applied computer vision and robotics-adjacent systems.
-
SARProcessing.jl- Built parts of a Julia pipeline for turning raw satellite radar data into usable image products
-
Landslide detection in Greenland
- Built a detection pipeline by faking training data, as I had exactly two real landslides to work with 🥹
- Mixed terrain geometry, radar decorrelation, area stats, and a fairly stubborn data pipeline until it worked
-
Glacier ice and sea ice from Landsat
- Authored research showing that glacier ice and sea ice near glacier fronts can be detected from ordinary Landsat imagery
- Despite what most expected, the segmentation idea I had was possible because changes in water content show up in the infrared bands.
SARProcessing.jleurosat-mlops-pipelinegraph-nnets-demo
- 🉐 Languages: Python · Julia · MATLAB · C++
- 🤖 ML: PyTorch · U-Nets · GNNs
- 👓 Computer vision: Segmentation · feature-based image analysis · OpenCV
- 🛠️ Tooling: GitHub Actions · Docker · DVC · remote GPU/HPC
- Combining webcam and a ultrasound distance measurement through the LeJEPA ideas to build depth aware pixels in a scare/noisy data environment.
- A landslide revival, where I am trying to turn an old research model into something closer to a usable monitoring pipeline