Nicolas Karasiak
Senior Geospatial ML Engineer
10+ years building large-scale geospatial ML systems and remote-sensing platforms: architecting full pipelines from raw data to customer product, distributed processing, time-series analysis, and open-source tooling.
600+ scientific papers built on my open-source geospatial software
300+ citations of my peer-reviewed research.
EXPERIENCE
EarthDaily, Data Scientist / ML Engineer Toulouse · September 2022 to Present
- Built a geospatial platform tracking hundreds of millions of farm parcels, including a custom zonal-statistics engine that is faster and lighter on memory than xarray.
- Cut processing costs ~100x and raised first-run success to 99% by redesigning the distributed ML pipelines around that engine, an ARM64 migration, and a custom storage format.
- Built a simulation of the EarthDaily constellation across sensors from low to very high resolution, covering temporal, spatial, and spectral dimensions.
- Fine-tuned geospatial foundation models (Tessera, Prithvi, Galileo, Presto) for crop classification and yield estimation.
- Subject-matter expert on Xarray, STAC, and the EarthDaily Python SDK. Defined how the team turns raw data into training sets.
Pixstart, Remote Sensing Data Scientist Toulouse · September 2020 to September 2022
- Built change detection workflows on Sentinel-2 time series for linear infrastructure monitoring
- Built ML models with expert rules and various geographic data
providers to produce ready-to-use map data
Dynafor / INRA, PhD, Remote Sensing Time Series Toulouse · October 2016 to August 2020
- Mapped forest species from satellite time series (Sentinel-2, Formosat-2) using phenology and canopy-structure features.
- Studied how spatial autocorrelation skews geospatial ML evaluation, and made the performance metrics more reliable.
- Published peer-reviewed research and shipped remote-sensing tools used in production.
Guiana Amazonian Park, Remote Sensing Intern Cayenne · March 2016 to September 2016
- Developed geospatial and remote-sensing pipelines using Python, R, QGIS, and OTB.
- Designed vegetation detection and monitoring workflows at park-wide scale using satellite imagery.
PROJECTS & OPEN SOURCE
- EarthDaily Python SDK: built the official package; adopted company-wide and extended after acquisition
- QGIS Dzetsaka Plugin: 200k+ downloads, used in 600+ scientific publications
- QGIS MCP Plugin: 100+ GitHub stars
- MuseoToolBox: Geospatial ML library with 43k+ downloads for reproducible remote-sensing workflows
SKILLS
- Languages
- Python, R, Rust (AI-assisted)
- Distributed & Performance
- Dask, distributed pipelines, large-scale geospatial processing, ARM64 profiling, bottleneck hunting, cost-efficient computing
- Geospatial
- STAC, GeoParquet, GDAL, Zarr, Xarray
- Machine Learning
- scikit-learn, PyTorch, Lightning, time-series
- AI Agents
- Claude Code (terminal agents), MCP servers, agentic workflows, prompt engineering
- Cloud & Infrastructure
- Linux (daily driver), AWS (EC2, EKS, S3), Docker, Argo
- Technical Direction
- Architecture, data platform & API/SDK design, ML pipelines, model selection, internal tooling, team mentoring
- Spoken Languages
- French (native), English (professional working proficiency)