TU Eindhoven · CS & Data Science

Systems that have to survive the real world.

I try to understand things properly: how a disease shows up in an image, how a machine can move like a person, how a structure fails under test. Then make something out of it worth putting into the world.

Click the field to plant something
01

Writing

RSS

Thinking out loud — about how we come to know things, and what that costs.

First posts in progress.

02

Projects

Things I have built, mostly with other people.

Professional work · Structural health monitoring

Sensor placement for fatigue monitoring

Where to put sensors on a large structure under test, so that damage is observable at all.

2026
High Tech Systems

MIMIC

Hardware that moves a real mouse, trained on human movement. Nothing runs on the host computer.

What it is

The host machine sees an ordinary mouse moving; everything else happens on the desk. Movement is learned from recorded human traces, so the paths look like a person's rather than generated ones.

Method
Vision
Screen state read externally, no host-side hooks.
Sequence
Sequence model over recorded human traces.
Actuation
Two degrees of freedom, calibrated to the desk.
It running
Where it lives
Repo private · on request
2026
Medical AI

Beyond Black Boxes

A chest X-ray classifier that shows where it looked and how sure it is.

What it is

Accuracy alone says nothing about a single case. Every prediction carries a confidence and the region it was read from, so the failures can be seen rather than averaged away.

Method
R-Net
Residual backbone over the radiograph.
Focal loss
The findings are rare and the classes badly unbalanced.
AdamW
Decoupled weight decay for fine-tuning.
2026
Crowd dynamics

Detecting phase-like states in pedestrian crowds

Whether a busy train platform can be described with physical phases. Mostly it cannot, and where it breaks is the interesting part.

What it is

A week of LiDAR trajectories from Eindhoven Centraal platform 2.1, cut into 3 m × 3 m cells over 5-second windows and clustered without labels. Of five features, only density and mean velocity order the way physical phases predict, and the states sit close together — real, but not cleanly separated.

Method
Filtering
Savitzky–Golay smoothing, then a physics filter for LiDAR identity switches and Kalman drift.
Cells
3 m × 3 m over 5 s, walkable floor only.
Clustering
Gaussian mixture, K = 4, on four z-scored features; flow excluded as density × speed.
A region slowing down
Detecting phase-like states in pedestrian crowds
2026
Autonomous Systems

Digital Twinning

A farm robot in simulation with a live twin beside it, mirroring its state as it drives.

What it is

A TurtleBot3 drives a simulated field while a command centre mirrors zone moisture, nutrients, battery and fertiliser. Weather and faults are injected from the twin to see what the robot does about them.

Method
ROS 2
Nodes for zone detection, crop decisions, resources, and the twin supervisor.
Gazebo · Nav2
Simulated field, SLAM, and navigation between zones.
Dashboard
Live telemetry, plus weather and fault injection from twin to robot.
Command centre, navigation, and the simulated field
Digital Twinning
Where it lives
2026
Software Design

FocusFlow

A Pomodoro tracker in Java that ties every session to a task.

What it is

A session cannot exist without a task, so the record afterwards says what the time went into, not just how much of it there was.

Method
JavaFX
Desktop UI.
Repository
Persistence behind one interface.
Facade
One entry point over the services.
Where it lives
03

About

I am a computer science and data science student at TU Eindhoven. The subjects change — robotics, medicine and disease, econometrics — but the pattern does not: learn one far enough to build something with it, then find out whether the thing survives contact with the world.

Elsewhere

The header is grown live in your browser. No two visits alike.