Giannis Dravilas
Work Experience
- Operational Forecast Validation: Designed and developed a benchmarking system to evaluate weather forecasts against 8,000+ real-time stations, utilising results to drive model selection.
- Hyperlocal Forecasting Pipeline: Designed and engineered a project to enhance NWP accuracy by assimilating localised station data, transitioning the system from PoC to a production microservice.
- Data Quality Control (QC): Optimised Python-based QC mechanisms for meteorological time-series; refactoring legacy code yielded a 50% speed improvement.
- Visualisation: Built interactive dashboards and visualisations to communicate complex weather concepts to stakeholders and the public.
- Mentorship: Mentored a junior data scientist on best practices for scientific software engineering.
- Python
- SQL
- NetCDF
- GRIB2
- Parquet
- xarray
- Docker
- Git
- S3
- AWS Athena
- Grafana
- Meteorological Data
- Numerical Weather Prediction
- Weather Station Observations
- Timeseries
- Machine Learning
- AI-Factory Data Workflows: Explored concepts for data management in support of establishing AI-Factories within the context of Destination Earth.
- Deep Learning on Satellite Data: Applied Vision Transformers (ViTs) and CNNs to MTG and OPERA remote sensing data, in cross-collaboration with the Remote Sensing Products Division.
- Cloud-Native ML Workflows: Built AI-ready EO datasets from MTG/OPERA data on the DestinE S3 Data Lake using cloud-native Zarr pipelines, resolving I/O bottlenecks.
- HPC Scalability: Scaled model training across multiple GPUs on HPC systems, utilising DDP.
- MLOps: Containerised ML workflows (Docker), automated inference with Airflow, and tracked experiments with MLflow.
- Python
- PyTorch
- Zarr
- HPC
- Distributed Data Parallel
- Docker
- Apache Airflow
- MLflow
- S3
- GitLab
- Deep Learning
- Vision Transformers
- Convolutional Neural Networks
- Remote Sensing
- Satellite Data
- Weather Radar Data
- Focused on the connectivity requirements and network design for High-Performance Computing (HPC) systems in Europe.
- Conducted in-depth analysis of unstructured text data, exploring trends and deriving actionable insights.
- Used visualisation techniques to effectively communicate key findings to stakeholders.
- Contributed as an independent contractor for the development of a Climate Risk Assessment Platform.
Education
- Thesis (in progress): ML-Based Feature Detection on GEO-Ring Satellite Data, in collaboration with EUMETSAT. Supervisors: Ana Lučić (UvA), Miruna Stoicescu (EUMETSAT), Roope Tervo (EUMETSAT).
- Research outputs: Two journal submissions in the area of Vision Transformers (1, 2), one submission in Information Retrieval (under review), one accepted paper in Human-Computer Interaction.
- Thesis: Machine Learning Snowfall Retrieval Algorithms for Satellite Precipitation Estimates. Supervisors: Manolis Koubarakis (NKUA), Stavros Dafis, Georgios Kyros, Kostas Lagouvardos (NOA).
- Thesis grade: 10/10 - selected for the annual e-book showcasing top theses from the Department, resulted in a peer-reviewed publication.
Publications & Research
-
Burning Lowlands: A Serious Game to Evaluate Citizen Learning, Communication, and Decision-Making in Climate Adaptation
EGU General Assembly, 2026. [DOI] -
InfoCIR: Multimedia Analysis for Composed Image Retrieval
arXiv [cs.HC], 2026. To appear in IEEE PacificVis 2026. [arXiv] -
On the Reproducibility of Vision Transformers Need Registers
Submitted to TMLR. [OpenReview] -
Reproducibility Study of FairViT: Fair Vision Transformer via Adaptive Masking
Submitted to TMLR. [OpenReview] -
Connectivity Requirements and Network Design for HPC Systems in Europe: Summary Report
Publications Office of the European Union, 2025. [DOI] -
Designing a Global Weather Station Network
EGU General Assembly, 2024. [DOI] -
Towards a Machine Learning Snowfall Retrieval Algorithm for GPM-IMERG
Environmental Sciences Proceedings, 2023. [DOI]
* Equal contribution
Projects
Built a web application with interactive 2D and 3D snowfall forecast maps for Europe, based on downscaled NWP models in 400 m spatial resolution.
Developed InfoCIR, a modular Plotly-Dash visual analytics dashboard that integrates multimodal composed image retrieval (CIR) models into an interactive prompt-engineering and debugging system.
Designed and implemented a tool to evaluate the accuracy of multiple weather models, using observational data from 8,000 WeatherXM weather stations.
Refactored and optimised all the quality of data mechanisms used at WeatherXM, resulting in a 50% speed improvement.
Developed a high-resolution, post-processing, rule-based snow forecasting model for Greece in Python using downscaling techniques.
Applied ViTs to large-scale weather data, used DDP on an HPC system, built Zarr datasets on S3 and employed Docker, MLflow and Apache Airflow.
Explored the role of register tokens in Vision Transformers by reproducing prior work, testing generalisation across diverse models (DINO, DINOv2, OpenCLIP, DeiT3).
Reproduced and evaluated the FairViT model's fairness and efficiency claims on two datasets, revealing limitations in adaptive masking.
Conducted Sentiment Classification of users' text reviews using Deep Learning (FFNN, RNN, and a pre-trained BERT model in PyTorch) and developed a QA engine.
Created a music genre classifier for short music clips using PyTorch-based Feed-forward and Convolutional Neural Networks.
Visualised weather data in Python, using GRIB and netCDF formats, as well as Matplotlib and Cartopy libraries.
Developed a waterspout forecasting model using the SWI Index and numerical weather data in Python.
Developed a facial recognition system using Eigenfaces (PCA + nearest neighbour) on the Yale B dataset, and a handwritten digit classifier using SVM.
Conducted Sentiment Analysis on Covid-19 tweets obtained from Twitter, using a range of Data Mining techniques and Machine Learning algorithms.
Analyzed data of New York City taxi trips and created a Random Forest Regressor for predicting the duration of taxi trips.
Solved the constraint satisfaction Radio Link Frequency Assignment Problem using Artificial Intelligence.
Implemented polygonisation algorithms for point sets in a plane, and optimised their areas using operational research concepts in C++ with the CGAL library.
Developed software and protocols in C/C++ for Linux, using processes, threads, pipes, signals and the internet protocol.
Developed parallel solutions for Monte Carlo simulations using Pthreads, OpenMP and MPI.
Volunteer & Extra-Curricular
Founder and owner of Weatherathens and snowly.eu. Nature photography work featured online by research institutions, e.g., National Observatory of Athens.
Translate user interface content and computer science terminology for the Indico project from English to Greek.
Led a hands-on workshop on data science and ML for weather/climate using Python and key libraries.
Supported first-year students in understanding core concepts and completing lab exercises for the courses "Introduction to Programming" and "Data Structures and Programming Techniques".
Translated user interface content and meteorological terminology from English to Greek.