ML engineer working on computer vision and GeoAI. Currently ML Lead at Dira GeoSystems in Zurich, where I build models that turn aerial imagery into maps, and the service that runs them.
Open-source contributions to cog, rules_python, bazel-watcher, langfuse-python, Building-Regulariser and RustGIS.
Posts
- Nov 11, 2026 · Testing a model that outputs polygons: how strict is strict enough?
- Sep 8, 2026 · Rollerblades and rocket ships: keeping an early inference pipeline ready for scale
- Aug 24, 2026 · “No kernel image is available”: catching torch and GPU mismatches before you ship
- Jul 8, 2026 · One Python version for Bazel: pyproject.toml as the single source of truth
Talks
- Jun 2026 · A good data model is worth a complex neural network. Oral presentation at GeoAI 2026, Ghent. Slides
Interviews
- Sep 2026 · Every training label was lying, and the loss curve looked fine. Interview with Ludovico Bessi for ML@SCALE. Read
Publications
- Zhou, Z., Winkler, J., Fu, C. and Weibel, R. (2026). A good data model is worth a complex neural network: evidence from building generalization using deep learning. Proceedings of the 1st International Conference on Geospatial Artificial Intelligence (GeoAI 2026). Paper