AI Lead & Computer Vision Engineer · Linköping, Sweden
I teach machines to read the forest from above. Before that, I spent fourteen years reading the earth beneath it.
One thread, three chapters
My work has always been about the same question: what is really in this signal, and can I trust it? Only the signal has changed, from seismic waves to drone images.
2005 – 2019
Seismic field operations, NISOC
Fourteen years on land seismic crews, from junior to senior QC supervisor: making sure every sensor recording was clean, consistent and in the right place before anyone built on it.
Carried forwardSignals, sensors, geometry, and a habit of checking data before trusting it.
2019 – 2021
MSc Statistics & Machine Learning, Linköping University
A second master's degree to turn field intuition into modelling skill, finished with a thesis on anomaly detection and causal inference in 5G networks together with Ericsson Research.
Carried forwardStatistics, machine learning, and causal thinking.
2022 – today
Arboair AB, Linköping
Computer vision for drone and aerial imagery of forests: I designed and lead the company's computer-vision platform and the models that map tree species and forest damage.
Carried forwardDeep learning, geospatial ML, and leading AI work end to end.
Now
At Arboair I lead the AI work behind forest analytics from drone and aerial imagery. Details stay with the company; this is the shape of it.
More about my workLearning notes
The latest note from each topic. Short, single-topic articles written from my own notebooks and experiments, spanning 2017 to 2026.
What a monotonic constraint is, how boosting libraries enforce it while growing trees, and when it makes a model more accurate and easier to trust.
Why random K-fold gives scores that are too good when samples come in groups or sit close together in space, and how splitting by group or by spatial block gives an honest estimate.
How the shoelace formula turns a list of corners into an area, why it works (trapezoids and cross products), what its sign tells you, and two traps: self-crossing rings and coordinates far from the origin.
Why DINOv3's patch features got worse the longer it trained, how the Gram anchoring loss fixes that, and what else changed since DINOv2.
How normalising rows and columns over and over turns a positive matrix into a doubly stochastic one, and why the same loop gives a soft, differentiable optimal transport plan (a soft version of the assignment problem).
What each common 3D format stores, how much memory it takes, how to convert between them, and when to use which.