Portrait of Gert Dekkers

Hi, I'm Gert.

Machine learning engineer at heart. I build AI that works outside the lab, from raw sensor signals to systems running in production.

I build AI that has to survive contact with the real world: on resource-constrained edge devices, in harsh conditions such as space, and in care homes where a false alarm or a missed fall actually matters.

My PhD at KU Leuven was about recognising daily activities with a wireless network of microphones — on an energy budget. At Magics Technologies I went from ML engineer on industrial ML projects to leading a team of chip designers and software/ML engineers building a space-grade neural network accelerator and software stack.

Today I'm a Senior AI Specialist at Nobi, working on vision-based intelligence helping people in need.

Experience

  1. Senior AI Specialist

    NobiAug 2025 – present · Belgium

    Working on vision-based intelligence for smart lights in elderly care. I have primarily been working on reducing the false-positive rate of our fall detection system.

    • Halved the false-positive rate of our fall detection system
  2. Team Lead AI Engineering

    Magics TechnologiesSep 2022 – Jun 2025 · Geel, Belgium

    I contributed to the development of a space-grade neural network accelerator, where reliability comes from radiation-hardening in both hardware and software. I led a heterogeneous team of digital chip designers and software/ML engineers, while at the same time contributing to the development of the software ecosystem. I co-supervised two PhD students where I initiated the project funding for one.

    • Resource management and performance evaluation
    • Technical lead for all software activities
    • Development of a software stack to deploy neural networks on custom accelerators
    • Project proposal writing
  3. Machine Learning Engineer

    Magics TechnologiesApr 2021 – Sep 2022 · Geel, Belgium

    I was responsible for anything from data science to project management on industrial ML problems. The main project was the development of an edge device along a central management system for the fleet capable of ultrasound AI inference with ultrasound as data modality.

    • Syringe break detection from ultrasound with neural networks (medical)
    • Regression of cracks/expansion in packaged nuclear waste from acoustic emission
  4. Research Intern

    SonosJun 2019 – Dec 2019 · Santa Barbara, USA

    Designed and evaluated neural network architectures for estimating room geometry from multi-channel acoustic impulse responses.

  5. PhD Researcher

    KU LeuvenJan 2015 – Apr 2022 · Geel, Belgium

    Automatic classification of daily activities at home with a wireless acoustic sensor network — from collecting the dataset, learning an ADL classification model, to making every node cheaper to run.

    Education PhD in Engineering Technology (2021)— thesisSound recognition with resource-constrained wireless sensor networks ↗

    • Built the SINS dataset
    • Dynamic sensor activation during late fusion
    • Joint quantisation and pruning for mixed-precision arithmetic
    • Organised DCASE 2018 Task 5
  6. Research Assistant

    KU LeuvenNov 2013 – Dec 2014 · Geel, Belgium

    Designed and implemented a multi-channel noise reduction algorithm for speech recognition.

  7. Education

    MSc Engineering Technology, Electronics-ICT

    KU Leuven2011 – 2013

    Cum laude · thesis on 2D audio source localisation with an acoustic sensor network

  8. Education

    Professional Bachelor Electronics-ICT

    KH Kempen2008 – 2011

    Cum laude

Skills

AI / ML
PyTorch (quantization), TensorFlow (Lite), Apache TVM, Xilinx Brevitas, ONNX, MLflow, Weights & Biases
Engineering
Embedded AI, Quantisation & pruning, (Multi-channel) signal processing, Sensor fusion
Sensing
Audio & acoustics, Vision (cameras), Ultrasound, Acoustic emission
Programming
Python, uv, Poetry, Ruff, pytest, C, Bash, LaTeX, Git, Claude Code, OpenCode
Infrastructure
Docker (compose), CMake, GitHub Actions, Grafana, Prometheus, PostgreSQL, AWS, Google Cloud
Leadership
Technical supervision, Mentoring, Team wellbeing, Project management, Project acquisition
Languages
Dutch, English

Online sources

  1. 2024
    Understanding the Limitations of Single-Event Upset Simulation in Deep Learning Libraries

    SPAICE 2024

  2. 2024
  3. 2022
  4. 2021
  5. 2021
    QliP: joint quantisation and pruning by iterative ranking of decomposed saliency-cost ratios

    Technical report · first author

  6. 2018
  7. 2018
  8. 2017
  9. 2015