Building (geo-)data pipelines and geoprocessing workflows — from ETL automation and data modeling to satellite-based damage assessment (caused by natural disasters or conflict) and environmental monitoring to tackle environmental crimes via satellite images. Combining 12+ years of hands-on geospatial expertise with modern ML engineering practices to solve real-world problems across agriculture, development cooperation, civil protection, and geointelligence applications.

ML & Data Science	MLOps & DevOps
Geospatial & Remote Sensing	Cloud & Visualization

How hard did the 2026 summer drought hit grassland in the Canton of St. Gallen? A reproducible Sentinel-2 pipeline, run from Python on Google Earth Engine, that measures grassland condition and mowing field by field in three landscapes, against a 2019–2025 baseline and MeteoSwiss station data.

Key Features:

• Pixel purity modelled from the Sentinel-2 point-spread function, so small, edge-dominated meadows are measured without mixed-pixel contamination
• Result: the meadows' moisture index (NDMI) fell 23–29 % below normal, the lowest of eight years at both valley sites
• Mowing detector validated like a model: rules frozen after tuning, then scored on 136 hand-labelled field-years including a hold-out (F1 0.87), with its drought-year failure diagnosed
• Runs end to end on calibrated synthetic data; provenance checks refuse stale results; CI with tests

Tech Stack: Python • Google Earth Engine • Sentinel-2 • GeoPandas • rasterio • NumPy • pandas • SciPy • pytest • GitHub Actions

🚧 Under active development · repository not yet public

End-to-end MLOps pipeline for automated building damage assessment from satellite imagery using YOLO26 / RT-DETR. Full ML lifecycle from data versioning to production deployment.

Key Features:

• FastAPI microservices architecture with JWT authentication, being prepared for Kubernetes deployment
• Automated CI/CD with model promotion logic
• Real-time monitoring and drift detection
• Production-ready API serving

Tech Stack: YOLO26 • RT-DETR • PyTorch • DVC • MLflow • GitHub Actions • FastAPI • Docker • Prometheus • Grafana • Evidently

🚧 Under active development.

Routed retrieval-augmented generation (RAG) system over the UCDP Georeferenced Event Dataset. Combines semantic search with deterministic aggregation on a fully local LLM stack — no data leaves the machine, and every answer stays traceable to the source events behind it.

Key Features:

• Hybrid BM25 + dense retrieval (BGE-large-en-v1.5) fused via reciprocal rank fusion, with multi-query expansion and BGE reranker
• Four-path query router separating semantic retrieval, deterministic aggregation, and out-of-scope refusal
• Source-bias surfacing and geospatial visualization of results
• Non-circular evaluation harness: method-independent gold set via 5-arm pooling, multi-event held-out recall benchmark, nDCG/MAP/MRR
• 214 unit tests with CI; Docker setup for reproducible local runs

Tech Stack: Python • ChromaDB • BM25Okapi • BGE-large-en-v1.5 • Mistral 7B via Ollama (local) • Streamlit • Docker

🚧 Under active development · repository not yet public

Machine learning system for detecting terrestrial oil spills using multi-temporal Sentinel-2 imagery. Handles extreme class imbalance and spectral confusion through temporal change detection.

Key Features:

• Multi-temporal baseline composites for change detection
• Custom augmentation for extreme data scarcity (only few ground validated reference events)
• MLOps pipeline from data ingestion to containerized model serving

Tech Stack: PyTorch • U-Net • Prithvi-EO-2.0 • Clay • Sentinel-2 • DVC • MLflow • BentoML • Docker • Microsoft Planetary Computer

A hands-on best-practice guide for annotating satellite and aerial imagery for object detection, paired with a curated collection of publicly available, high-quality annotated remote sensing datasets.

Key Features:

• Annotation best-practice guide: bounding boxes & OBB, model-assisted labeling (SAM 2/3), GIS-based tiling workflows, occlusion & labeling conventions, quality assurance, and spatially blocked data splits — backed by references to current literature
• Overview and categorization of datasets by application (building footprint extraction, road segmentation, ship detection)
• Direct links to data sources and relevant academic papers, with clear inclusion criteria focusing on data quality, documentation, and accessibility

• Current Focus: Deep Learning & Remote Sensing, GIS & Data Science Specialist
• Previous: Remote Sensing & ML Specialist (10+ years)
• Education: PhD in Environmental Remote Sensing & Machine Learning
• Research: 10+ peer-reviewed publications in GIScience & Remote Sensing, Environmental Management, and Applied Remote Sensing