Skip to content
MDM

Magnus De Maerteleire, AI engineer

I build AI systems, from research models to agents I use every day.

MSc in Artificial Intelligence from KU Leuven, cum laude. Now at INSEAD for the Master in Management, in France and Singapore.

For my thesis I tested genomic deep-learning models on 190 personal genomes. With three classmates I developed a CLIP labelling method that beat a published baseline on three benchmarks, and I built a sandboxed multi-agent system that runs on my own server.

During my bachelor's I worked as a developer at Lux49, a medical software startup, where I built an AI model that measures muscle activity from facial images. Alongside my studies I also founded FCO2, a motorsport sponsorship company in Luxembourg.

Available
July 2027
Work rights
EU citizen
Currently
INSEAD, France and Singapore

Index8 projects

ProjectKindKey resultYear
Predicting regulatory variation from personal genomes2026
Neurosymbolic AI for handwritten formulas2026
Automatic concept labels with CLIP2025
Sandboxed multi-agent assistant
Active-learning ranking app2026
Constrained translation with grid beam search2025
Near-duplicate detection in 5 million tweets2025
NBA player tracking with Spark2025

Selected projects4 of 8

  • Predicting regulatory variation from personal genomes

    2025–2026 / MSc thesis, Aerts Lab, KU Leuven / Code private (lab data)

    Can deep-learning models predict how one person's DNA changes gene activity, and does that help predict Parkinson's disease? A large-scale test on 190 personal genomes.

    Spearman up to 0.89 at cohort level, but below r = 0.10 per donor: a clear, quantified limit of today's models.

    Correlation with measured data, cohort level versus donor levelCohort level: Scooby, gene expression, substantia nigra: 0.89; CREsted, chromatin, substantia nigra: 0.79; CREsted, chromatin, cingulate cortex: 0.73. Donor level: CREsted, chromatin, substantia nigra: 0.10; CREsted, chromatin, cingulate cortex: 0.09; Scooby, gene expression, substantia nigra: 0.0600.250.50.751Cohort levelSpearman ρ, prediction vs cohort-average measurementScooby, gene expression, substantia nigra: 0.885Scooby gene expression, substantia nigra0.89CREsted, chromatin, substantia nigra: 0.792CREsted chromatin, substantia nigra0.79CREsted, chromatin, cingulate cortex: 0.734CREsted chromatin, cingulate cortex0.73Donor levelPearson r, best cell type, predicted vs measured deviation per donorCREsted, chromatin, substantia nigra: 0.095CREsted chromatin, substantia nigra0.10CREsted, chromatin, cingulate cortex: 0.093CREsted chromatin, cingulate cortex0.09Scooby, gene expression, substantia nigra: 0.063Scooby gene expression, substantia nigra0.06
    Both models get the cohort average right. Per donor, even the best cell type stays below r = 0.10, and only on regions pre-selected for known regulatory effects; on the Parkinson’s regions the signal is indistinguishable from zero.
    • Python
    • PyTorch
    • Scooby
    • CREsted
    • scikit-learn
    • HPC
    • Single-cell multiome
    Read the write-up
  • Neurosymbolic AI for handwritten formulas

    2026 / Learning and Reasoning course, KU Leuven / Code on GitHub

    Reading a sequence of handwritten digits and operators and computing the result, comparing logic built into the model with logic added to the loss.

    With five images, a plain neural network falls to 7.1% accuracy; putting the logic in the architecture keeps it at 98.7%.

    3 images: Neural baseline 86%; Logic in the architecture 97.8%; Logic in the loss 96.5%. 5 images: Neural baseline 7.1%; Logic in the architecture 98.7%; Logic in the loss 28.9%3 imagese.g. 7 + 2, mean of 3 runsNeural baseline86.0%Logic in the architecture97.8%Logic in the loss96.5%5 imagese.g. 7 + 2 − 4, one runNeural baseline7.1%Logic in the architecture98.7%Logic in the loss28.9%
    Test accuracy on the final result. With five images the plain network has to learn too many combinations; with the rules in the architecture it only has to recognise single symbols.
    • Python
    • PyTorch Lightning
    • Probabilistic logic programming
    • CNNs
    Read the write-up
  • Automatic concept labels with CLIP

    2025 / Project AI course, KU Leuven, team of four / Code on GitHub

    Labelling the concepts that explainable AI models need, without any manual labelling.

    Beat the Stochastic-CBM baseline on all three datasets, from 73.1% to 84.1% balanced accuracy on Pascal VOC.

    1. Step 1

      Similarity

      CLIP embeds every image x and concept description t. Their cosine similarities fill the matrix: sij = EI(xi)⊤ ET(tj).

    2. Step 2

      Two Gaussians

      For each concept (one column), two Gaussians are fitted to the histogram of similarities.

    3. Step 3

      Ranking

      A custom score measures how well the curves separate. Poorly separated concepts are removed.

    4. Step 4

      Thresholds

      Each kept concept gets its own threshold, computed from its fitted Gaussians.

    The four steps. Only step 1 uses the images; the labels are used to evaluate, never by the method.
    • Python
    • PyTorch
    • CLIP
    • SciPy
    • NumPy
    Read the write-up
  • Sandboxed multi-agent assistant

    Personal project / In daily use, code private

    A multi-agent system on my own server that keeps my planning on track, from INSEAD deadlines to my job search. Agents that research the open web run inside a sandbox; the assistant that works with my personal accounts acts behind explicit approval rules.

    In daily use around the clock: my planning, deadlines and 200+ researched roles in one place; sending, submitting or deleting always waits for my approval.

    Architecture of the personal agent systemI talk to the assistant agent from my phone. It works with my personal accounts through connectors, within fixed permissions. The sandboxed research agent scans the open web. The agents write to one job database, and the ranking app reads from it. Everything runs on my own server.Me, on my phonetext and voice notesMy accountsconnectors, fixed permissionsOpen webcareer sites, job postsAssistant agentplanning, deadlines, remindersResearch agentsandboxed: no accounts or filesJob databasea permanent id for every roleRanking appmy devices onlydaily scanwriteswritesMy own server, always on.Scheduled work runs on the server. Memory in files under version control.
    The core agents, one job database, one server; more agents can be added for new tasks. The research agent never touches my accounts or files.
    • Claude Code
    • MCP
    • Python
    • SQLite
    • Linux
    • Local speech-to-text
    Read the write-up

Experience

  1. FCO2

    Luxembourg

    Motorsport sponsorship, events and vehicle wrapping

    Managing Director

    2022–now

    Founded the business next to my studies: 14 racing sponsorship deals across five series, a vehicle-wrapping line I started from scratch, and a team of three.

    See the cars and deals
  2. Howald, Luxembourg

    Medical software startup

    Developer

    2021–2022

    Built an AI model that measures muscle activity from facial images, and front and back ends for medical apps across the Apple ecosystem, in a five-person team.

    Visit lux49.lu

Education

  1. INSEAD

    France and Singapore

    Master in Management

    2026–2028

    VP of Careers at the Tech, Media & Telecom club; member of the AI and Entrepreneurship clubs. Graduating January 2028.

  2. KU Leuven

    Leuven, Belgium

    MSc Engineering: Artificial Intelligence

    2024–2026

    Graduated cum laude. Thesis on deep-learning models that predict gene regulation from personal genomes, applied to Parkinson’s disease. Student representative for the first cohort.

    Read about the thesis
  3. KU Leuven

    Leuven, Belgium

    Bachelor of Engineering

    2021–2024

    Major in electrical engineering, minor in computer science.

Additional information

Work rights
Belgian citizen, EU
Languages
Dutch (native), English (fluent), French (conversational)
Technical
Python, PyTorch, Keras, C, SQL, JavaScript, Java, Next.js, AI-assisted development with Claude Code and Codex
Making
Posters and illustration in Illustrator and Photoshop; 3D printing, large-format printing, laser and vinyl cutting
Interests
Padel, chess, racing history, architecture