R&D Engineer at VITO Remote Sensing

Ioannis Kalfas

Senior Computer Vision Engineer: from research to production

I build deep vision models and take them from research prototype to operational deployment: global 10 m land-cover maps for Copernicus, insect-recognition cameras in the field, and compact models that compress satellite imagery on the edge.

View CV (PDF) Get in touch
  • Leuven, Belgium
  • Greek · EU national
  • PhD, KU Leuven
GitHub LinkedIn Google Scholar
Ioannis Kalfas presenting satellite-based flood mapping work on stage

On stage at ECMWF Code for Earth 2026

11yrs
in computer vision and deep learning, from visual neuroscience to production
9
peer-reviewed publications, 5 as first author, 6 on computer or biological vision
10m
resolution global land-cover maps for Copernicus LCFM
8+
master’s and bachelor’s theses supervised

About

One thread: teaching machines to see

I’m Ioannis (Yannis) Kalfas, a computer vision engineer and PhD with 11 years of experience building deep vision models and the systems around them. My work spans image classification, object detection, semantic segmentation and neural image compression, together with the data engineering, MLOps and optimization needed to run them in the cloud, on the edge and on embedded hardware.

I started in vision science: at KU Leuven’s neurophysiology lab I compared deep convolutional networks with how neurons in the primate visual cortex represent shape, and used CNN features to predict neuronal responses to images. I then applied vision to industry, to field imaging systems that recognise insects for my PhD and postdoc, and since 2024 to satellite imagery at VITO, first for global land-cover and cloud segmentation for Copernicus and now for efficient compression models on the edge.

I’m not only a researcher: I care about deploying what I build. Across my roles I’ve owned the path from model to running system, with Docker, cloud APIs, Git workflows, CI and MLOps tooling, from client projects at Faktion to field devices in my PhD and global production pipelines at VITO. Earth observation, agri-food and industry are the domains where I apply vision, and alongside images I work with time-series and sensor signals, from optical insect wingbeats to predictive maintenance.

  • Vision across scales

    From neurons and lab images to sticky-plate photos in the field and 10 m satellite imagery of the whole planet.

  • Efficient & edge vision

    Neural image compression (VQ-VAE) and models sized for Raspberry Pi devices and onboard hardware.

  • Ships to production

    Global EO products, an AWS-hosted model-serving platform, Raspberry Pi field devices, Dockerized releases and reproducible MLOps.

Career at a glance

Computer vision throughout: from neurons, to insects, to satellites.

  • Biological vision
  • Industrial vision
  • Field & agri-food imaging
  • Satellite imagery

From model to production, at every step

Deployment, MLOps, DevOps and Git workflows have been part of the job since my first industry role.

  • Faktion2017–18

    End-to-end ML for industrial clients, with standardized feature-extraction and data-preparation workflows reused across projects; a video activity-recognition prototype that won an industry hackathon.

  • PhD, KU Leuven2018–22

    Deep-learning models running on Raspberry Pi field devices (PhotoBox, flying-insect trap), connected to an AWS-hosted API built with FastAPI and Docker for external research collaborators.

  • Postdoc, KU Leuven2022–24

    Built and ran an AWS-hosted image-classification, annotation and model-serving platform used by external companies; led the move to shared Git infrastructure and engineering practices across research teams; managed multi-terabyte datasets on ManGO.

  • VITO2024–

    Operationalized global land-cover models: inference across local, AWS and openEO, Spark and Hadoop pipelines, MLflow integration, package maintenance and output validation. Now building MLOps tools for experiment tracking and GPU-cluster jobs.

  • Open source & side projectsongoing

    selfhostit: signed multi-arch Docker images, automated releases, a one-line installer with auto-updates, nightly backups and CI security scans. desto: a dashboard and CLI for running and scheduling jobs.

  • ECMWF Code for Earth2026

    Project Atlantis: a packaged pipeline with a CLI, a Python API and a consolidated Zarr/STAC archive for ML-ready flood data.

Experience

Eleven years of computer vision, in research and industry

R&D Engineer, Computer Vision for Earth Observation

VITO · Remote Sensing unit · Mol, Belgium

Mar 2024 – Present

One continuous role in VITO’s Remote Sensing unit, building deep vision models for satellite imagery and the pipelines that run them at planetary scale. The work has moved from segmentation and classification for global land-cover mapping to efficient neural compression for edge and onboard hardware.

Efficient vision on the edge: neural image compression

Current GeoAI and Edge Solutions team Jan 2026 – Present
  • Optimize and evaluate CORSA VQ-VAE architectures for satellite-image compression, balancing model footprint with spatial and spectral reconstruction quality on resource-constrained edge and embedded platforms.
  • Build parametrized training workflows with PyTorch, MLflow, Hydra and Pydantic to improve experiment reproducibility and model traceability.
  • Develop internal MLOps tools for experiment tracking, reproducible configuration, and job management on VITO’s on-premises GPU cluster.
  • PyTorch
  • VQ-VAE
  • Neural compression
  • MLflow
  • Hydra
  • GPU cluster

Global land cover: cloud segmentation & classification

ESA WorldCover & Copernicus LCFM team Mar 2024 – Jan 2026
  • Developed deep-learning models for cloud segmentation and global 10 m land-cover mapping from Sentinel imagery within the Copernicus Land Cover and Forest Monitoring programme.
  • Built multi-stage pipelines for satellite-data ingestion, preprocessing, model inference, and generation of annual global Earth-observation products.
  • Implemented reusable cloud-model inference workflows across local, AWS and openEO environments.
  • Engineered distributed data pipelines using Hadoop, Spark, object storage and network storage for geographically and seasonally diverse satellite datasets.
  • Improved model operationalization through MLflow integration, package maintenance, output validation, and standardized GeoTIFF metadata.
  • Supported technical handover by guiding colleagues through model configurations, datasets, annotation sources, training code and inference workflows.
  • Semantic segmentation
  • Image classification
  • Sentinel data
  • Spark
  • Hadoop
  • AWS
  • openEO

LCFM programme ESA WorldCover

PhD Researcher → Postdoctoral Researcher

KU Leuven · MeBioS, Biosystems · Leuven, Belgium

Dec 2018 – Mar 2024

Postdoctoral Researcher

Nov 2022 – Mar 2024
  • Built and deployed an AWS-hosted image-classification, annotation and model-serving platform used by research teams and external collaborators.
  • Led the adoption of shared Git infrastructure and software-engineering practices across multiple research teams.
  • Organized bimonthly deep-learning meetups for researchers across departments, communicating technical concepts to interdisciplinary audiences.
  • Managed multi-terabyte research datasets through KU Leuven’s ManGO platform, improving dataset organization and accessibility.
  • Investigated hyperspectral imaging and machine-learning methods for early disease detection in agri-food applications.

Demo: device setup Demo: web interface Demo: labelling tool

PhD Researcher in Optical Insect Identification

Dec 2018 – Nov 2022
  • Built image-based detection and classification of insects on sticky-plate photos taken in the field and by a Raspberry Pi imaging station (PhotoBox).
  • Designed multimodal datasets and training pipelines for optical and acoustic insect identification, including CNNs on optical wingbeat time series.
  • Developed deep-learning models for Raspberry Pi devices and integrated them with an AWS-hosted API (FastAPI, Docker) serving external research collaborators.
  • Supervised more than eight master’s and bachelor’s thesis students in computer vision and signal processing.
  • Published five peer-reviewed papers on computer vision and AI for biological monitoring and quality control.
  • PyTorch
  • CNNs
  • YOLO
  • Object detection
  • Time-series
  • Hyperspectral
  • Raspberry Pi
  • AWS
  • FastAPI
  • Docker

Data Scientist / AI Engineer

Faktion · Antwerp, Belgium

Dec 2017 – Dec 2018

  • Developed end-to-end computer-vision solutions and proofs of concept for industrial clients.
  • Won an industry hackathon by developing a video-based activity-recognition prototype.
  • Applied signal processing and time-series models to predictive-maintenance and speech-processing use cases.
  • Standardized reusable feature-extraction and data-preparation workflows for heterogeneous client datasets, improving consistency and reuse across projects.
  • Computer vision
  • Video analysis
  • OpenCV
  • Signal processing
  • Predictive maintenance
  • scikit-learn

Hackathon video

Doctoral Researcher in Vision Neuroscience & Deep Learning

KU Leuven · Neurophysiology · Leuven, Belgium

Nov 2015 – Nov 2017

  • Studied deep convolutional networks as models of the primate visual system, comparing CNN representations of shape with responses of macaque inferotemporal neurons and with human judgments.
  • Developed computer-vision and regression models that predict neuronal responses to visual stimuli from CNN features.
  • Analysed the shape selectivity of neurons in a body-selective region of the temporal cortex, and contributed the deep-network analyses to a study comparing DNNs with the human ventral visual pathway.
  • Published four peer-reviewed papers, including two as first author, and presented at the Vision Sciences Society conference before leaving the programme to pursue applied computer vision in industry.
  • Deep CNNs
  • Object & shape recognition
  • Neural encoding models
  • Regression
  • Python

VSS abstract

Selected work

Vision systems, from orbit to the field

CORSA: neural satellite-image compression on the edge

VITO · GeoAI and Edge Solutions · 2026–

Optimizing VITO’s lightweight VQ-VAE compression model so satellite imagery can be compressed on edge and embedded hardware, trading model footprint against spatial and spectral reconstruction quality. Alongside it, I build the MLOps tooling for experiment tracking and GPU-cluster job management.

  1. Satellite images
  2. CORSA compression (edge)
  3. On-prem GPU cluster

CORSA on Terrascope

LCFM: cloud segmentation & global 10 m land cover

VITO · Copernicus programme (€11M) · 2024–2026

Multi-stage AI pipelines producing annual global land-cover maps at 10 m resolution, ten times sharper than before, in support of EU environmental policy and climate monitoring. I focused on the cloud-segmentation models that identify usable observations and on deploying the classification models.

  1. Satellite data
  2. AI cloud detection
  3. Quality composites
  4. Global land-cover maps

LCFM ESA WorldCover (predecessor)

Computer vision for insect monitoring · KU Leuven

Imaging devices, vision and signal models, and an AWS-hosted platform that served external companies through a web interface.

PhD thesis
PhotoBox: Raspberry Pi, camera and LED lights imaging a sticky plate

PhotoBox

Raspberry Pi imaging station that photographs sticky plates for pest-insect classification.

 StickyBugs-AI illustration with insects on sticky plates

StickyBugs-AI

Detection and classification of insects caught on sticky plates in the field.

 Wingbeat frequencies illustration

Wingbeat AI

Insect species identification from optical wingbeat signals using CNNs.

 Flying-insect trap illustration

Flying-insect trap

Software for an automated trap that identifies flying insects in the field.

Project Atlantis

ECMWF Code for Earth 2026 · with Stylianos Lagaras · May – Aug 2026

A scalable STAC/Zarr pipeline for ML-ready, multi-source flood inundation observations, built during ECMWF Code for Earth, an innovation programme where developer teams work with ECMWF mentors on open-source Earth-science software.

It harmonises flood observations from VIIRS and MODIS (optical) and GFM (Sentinel-1 SAR) onto a common 1-arcmin grid, with a CLI, a Python API and a consolidated archive. The repository includes a worked example on the 2024 Valencia floods.

Mentors

Stack

  • Zarr
  • STAC
  • xarray
  • SAR
  • Optical EO
  1. VIIRS + MODIS + GFM
  2. Harmonised 1-arcmin grid
  3. Zarr / STAC archive
  4. ML-ready flood data

opageo/atlantis Code for Earth

selfhostit: small apps you run at home

Personal open-source project · creator & maintainer · 2026–

Outside my day job, I design, build and ship selfhostit, a growing family of free, MIT-licensed apps for everyday household life that run on your own hardware, from a Raspberry Pi to a NAS. No subscriptions, no cloud accounts and no tracking: each app keeps its data in one folder at home.

Every app grows from plasto, an open template I wrote with sign-in, a database, realtime sync, push notifications, scheduled jobs, nightly backups and a phone-installable web app already in place. A one-line installer sets up Docker, installs pinned and checksummed releases and keeps them updated. The apps run unprivileged on a read-only filesystem, behind Tailscale or a reverse proxy, with CodeQL, govulncheck, Grype and gitleaks scans and signed build attestations.

The apps

  • Splitwiz · split shared costs with a partner or roommate
  • should-we · weigh options together and get one ranking
  • Flowy · private period and cycle tracker, works offline
  • Troye · calorie tracker; food photos read by a vision model on your own GPU
  • Shoppa · shared shopping list with barcode scanning and Open Food Facts
  • Calio · one household calendar for plans, pet sitting and trips

Stack

  • Go
  • PocketBase
  • Svelte 5
  • Tailwind
  • PWA
  • Docker
  • amd64 · arm64 · armv7
  1. plasto template
  2. Household app
  3. Signed multi-arch images
  4. One-line install at home

selfhostit.org selfhostit selfhostit/plasto

Open source

Tools I build and maintain

Python tools that came out of day-to-day research and engineering work.

desto 100+

Web dashboard and CLI for running and managing Python and shell scripts in tmux sessions.

  • Real-time system monitoring and live log viewing
  • Script scheduling from the browser or the terminal
  • Full-stack: NiceGUI front end, Docker deployment
  • NiceGUI
  • tmux
  • CLI
  • Docker
vresto 29

Copernicus Sentinel product browser: interactive map search, OData/STAC backends, band visualisation, plus a CLI and Python API.

  • Copernicus
  • STAC
  • Sentinel-2
plakakia 12

High-performance image tiling for object-detection and segmentation datasets, using multiprocessing and NumPy.

  • Computer vision
  • NumPy
filoma 1

Profiles files, directories and image data to organise and validate large research datasets before ML pre-processing.

  • Data profiling
  • Validation
Home_Surveillance 11

Motion detection with OpenCV on a Raspberry Pi, with Flask streaming and push-notification alerts.

  • OpenCV
  • Raspberry Pi
  • Flask

Skills

Technical toolkit

Computer vision

  • Semantic segmentation
  • Image classification
  • Object detection (YOLO)
  • Neural image compression
  • Video activity recognition
  • Multispectral & hyperspectral
  • Deep CNNs
  • OpenCV
  • Image tiling & annotation

Earth observation

  • Satellite imagery
  • Sentinel data
  • Cloud screening
  • Land-cover mapping
  • SAR & optical EO
  • openEO
  • GeoTIFF
  • Zarr / STAC

Edge & efficient AI

  • VQ-VAE
  • Model optimization
  • Constrained inference
  • Raspberry Pi
  • Embedded deployment

Signals & time series

  • Time-series classification
  • Optical sensor signals
  • Predictive maintenance
  • Signal processing
  • Speech & audio
  • Neural encoding models

MLOps & data engineering

  • PyTorch
  • MLflow
  • Experiment tracking
  • Model validation
  • Hydra
  • Pydantic
  • PyTorch Lightning
  • scikit-learn
  • TensorFlow / Keras
  • Spark
  • Hadoop
  • xarray

DevOps, cloud & languages

  • Docker
  • Git workflows
  • AWS EC2 / S3
  • CI/CD & releases
  • FastAPI
  • Linux & Bash
  • On-prem GPU clusters
  • Self-hosting
  • Streamlit
  • NiceGUI
  • Go
  • Svelte
  • Greeknative
  • Englishprofessional
  • Dutch · Spanish · Swedishelementary

Publications

Nine peer-reviewed papers

Six of them on computer or biological vision, the rest on sensor signals and neuroimaging. Full list on Google Scholar.

  1. 2023
    Towards automatic insect monitoring on witloof chicory fields using sticky plate image analysis

    Kalfas I., De Ketelaere B., Bunkens K., Saeys W.

    Ecological Informatics 75, 102037First author

  2. 2023
    An introduction to artificial intelligence in machine vision for postharvest detection of disorders in horticultural products

    Tempelaere A., De Ketelaere B., He J., Kalfas I., Pieters M., Saeys W., et al.

    Postharvest Biology and Technology 206, 112576

  3. 2022
    Optical identification of fruitfly species based on their wingbeats using convolutional neural networks

    Kalfas I., De Ketelaere B., Beliën T., Saeys W.

    Frontiers in Plant Science 13, 812506First author

  4. 2022
    A fresh look at computer vision for industrial quality control

    De Ketelaere B., Wouters N., Kalfas I., Van Belleghem R., Saeys W.

    Quality Engineering 34(1), 152–158

  5. 2021
    Towards in-field insect monitoring based on wingbeat signals: the importance of practice oriented validation strategies

    Kalfas I., De Ketelaere B., Saeys W.

    Computers and Electronics in Agriculture 180, 105849First author

  6. 2019
    The ventral visual pathway represents animal appearance over animacy, unlike human behavior and deep neural networks

    Bracci S., Ritchie J.B., Kalfas I., Op de Beeck H.P.

    The Journal of Neuroscience 39(33), 6513–6525

  7. 2018
  8. 2017
    Shape selectivity of middle superior temporal sulcus body patch neurons

    Kalfas I., Kumar S., Vogels R.

    eNeuro 4(3)First author

  9. 2017
    Representation of semantic similarity in the left intraparietal sulcus: functional magnetic resonance imaging evidence

    Neyens V., Bruffaerts R., Liuzzi A.G., Kalfas I., Peeters R., Keuleers E., et al.

    Frontiers in Human Neuroscience 11, 402

PhD thesis · KU Leuven, 2022

Optical insect identification using artificial intelligence

MSc thesis · KTH, 2015

Simulating neocortical structures with spiking neural networks (NEST simulator)

Education

Three countries, three degrees

  1. 2018 – 2022

    PhD, Bioscience Engineering

    KU Leuven, Belgium

    Optical insect identification with signal processing, computer vision and deep learning.

  2. 2015

    MSc, Machine Learning

    KTH Royal Institute of Technology, Stockholm, Sweden

    Specialised in computational neuroscience and spiking neural networks.

  3. 2013

    BSc, Computer Science

    Aristotle University of Thessaloniki, Greece

    Foundations in computing theory and information systems.

Leadership & community

Beyond the code

  • Mentoring

    Supervised 8+ master’s and bachelor’s theses in computer vision and signal processing.

  • Deep-learning meetups

    Organised bimonthly meetups for researchers across KU Leuven departments.

  • Engineering practices

    Led the adoption of shared Git infrastructure and software-engineering practices across research teams.

  • Speaking

    Vision Sciences Society (Florida, USA) and ECMWF Code for Earth 2026.

  • Hackathon winner

    Video-based activity recognition prototype at an industry hackathon.

  • Open-source maintainer

    desto, vresto, plakakia and filoma, tools in active research use.

Contact

Let’s talk

I’m always interested in how computer vision can improve our lives, especially for Earth observation, efficient models on the edge, and turning research into dependable systems. Email is the best way to reach me.

Profiles

GitHub LinkedIn Bluesky Linktree

Research

Google Scholar ORCID ResearchGate

Résumé

CV (PDF, 2 pages)