Data Scientist | AI & Machine Learning | Weather AI, Climate & Earth Observation
Stylianos (Stelios) Papargyris
Thessaloniki, Greece
Available — Open to new roles
Data Scientist at the intersection of AI, weather, earth observation, and environmental data. I design and build systems that turn complex geospatial and atmospheric data into decisions — from ECMWF ensemble post-processing to real-time precipitation nowcasting across Europe. Currently seeking roles in applied ML, weather AI, or climate tech.
Precipiteau
Owner
A fully self-built, operational European precipitation platform covering 35+ countries, running 24/7. Delivers five distinct data products on a single interactive map: live radar rain-rate (EUMETNET OPERA), radar–satellite fusion (H-SAF MSG/SEVIRI fill-in), daily accumulation, raw reflectivity (ZhMax in dBZ), and a 1-hour optical-flow nowcast — every layer auto-refreshed every 15 minutes.
Custom-implements the Ramer ice-fraction algorithm against DWD ICON-EU vertical profiles downscaled to a 1 km terrain DEM, diagnosing precipitation phase (rain, snow, freezing rain, ice pellets, hail) across the full European domain. GPU-accelerated rendering with adaptive preloading for smooth animation on any device.
Open platform →- Designed and deployed a production XGBoost pipeline that bias-corrects ECMWF 51-member seasonal ensemble forecasts against ERA5 reanalysis across 14+ global regions — achieving +52% HSS improvement for precipitation and 30–50% RMSE reduction for temperature, with RMSE reductions of 25–29% across wind speed, humidity, and pressure; data ingested via Copernicus CDS and outputs delivered to AWS S3.
- Designed and deployed an automated parametric insurance resolution system for ParahubXM (WeatherXM's parametric insurance platform), processing ERA5 climate reanalysis to classify risk and resolve contracts across travel insurance and crop/farmer protection use cases via threshold-based precipitation analysis.
- Contributed feature additions and bug fixes to the Quality of Data (QoD) pipeline, including integrating ISD model scores into the QoD scoring logic and resolving edge cases in the MLflow and batch inference pipelines.
- Built a comprehensive automated pytest suite for the Quality of Forecasts (QoF) pipeline, covering forecast evaluation, meteorology utilities, observation processing, and data transformer modules — with fixture datasets for reproducible testing.
- Developed a reinforcement learning optimisation system for multi-source heat pump control, dynamically selecting between ambient air, geothermal, and solar energy sources to maximise COP — part of the SEEDS project.
- Designed edge computing solutions for IoT data execution and sharing in smart building energy management systems — part of the SEED4AI project.
- Prototyped ML models for predictive maintenance of photovoltaic systems, focused on State of Charge degradation forecasting.
Precipiteau Owner
Fully self-built operational platform delivering five precipitation data products across 35+ European countries: radar rain-rate, radar–satellite fusion, daily accumulation, raw reflectivity (ZhMax), and a 1-hour optical-flow nowcast. Custom Ramer ice-fraction algorithm on a 1 km terrain DEM for full precipitation-type diagnosis. GPU-accelerated rendering, 15-minute refresh cycle, 24/7.
ECMWF Bias Correction Pipeline
WeatherXM
Production XGBoost pipeline correcting systematic biases in ECMWF 51-member seasonal ensemble forecasts using ERA5 reanalysis across 14+ global regions — achieving +52% HSS improvement for precipitation and 30–50% RMSE reduction for temperature, with 25–29% RMSE reductions across wind speed, humidity, and pressure.
Environmental Pollution Detection using Deep Learning
MSc Thesis · Aristotle University of Thessaloniki
End-to-end deep learning pipeline for detecting air pollutants from silicon photonic SHFT spectrometer data — combining a denoising autoencoder for sensor compensation with a 1D CNN classifier. Research resulted in a paper submitted to EANN 2026.
Parametric Insurance Automation
WeatherXM · ParahubXM
Automated parametric contract resolution for travel insurance and crop/farmer protection, processing ERA5 precipitation fields to classify risk via threshold-based event triggers — fully replacing manual adjudication.
Weather Data Downloader
WeatherXM
Python CLI and library for bulk-downloading ERA5 reanalysis from Google Cloud ARCO-ERA5 (Zarr) and ECMWF seasonal forecasts from Copernicus CDS. Supports batch multi-station downloads with 3×3 grid spatial extraction, automatic unit conversion, geocoding, and structured Parquet output to local or S3 storage.
Coursework: Machine Learning · Deep Learning · Computational Intelligence · Biosignal Analysis · Intelligent Agent Systems
Coursework: Probability & Statistics · Signals & Systems · Algorithms & Data Structures · Control Systems · Renewable Energy · Sensors · Scientific Computing