Selected work
Benchmarking time-series foundation models for gyrokinetic plasma heat-flux prediction
MSc thesis in progress, expected 2026. Johannes Kepler University Linz, supervised by Fabian Paischer.
Turbulent heat flux in a tokamak is a single scalar, but obtaining it normally means integrating the 5D distribution function of a nonlinear gyrokinetic simulation, which is expensive. I am testing how close fine-tuning and in-context learning on time-series foundation models, given the scalar flux time series alone, can get to the accuracy of the GyroSwin-1B 5D neural surrogate, without modelling the 5D distribution function the way GyroSwin does.
Benchmarked against Paischer et al., GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations, NeurIPS 2025. arXiv:2510.07314
Apollonia, handwriting recognition for dental records
Contract project for AKS-ZAVOMED, an Austrian NGO, 2025 to 2026.
The project covers around 50,000 scanned handwritten dental examination forms. I trained a custom CNN, roughly 120,000 to 150,000 parameters over 12 classes on 32x32 grayscale regions of interest, to read per-tooth codes, reaching 96.93% per-cell accuracy on 3,257 of 3,360 cells. That took an automated experiment loop over 19 or more variants, plus around 1,200 hand-labelled cells to close the gap between public handwriting datasets (EMNIST, ARDIS, DIDA) and Austrian digit conventions. It runs fully on device to satisfy GDPR: ONNX INT8 export for CPU inference on an Apple M1 Max, with a PySide6 review interface for the operator.
Brainchair, open-source MCP server
Go, 2026.
An always-on Model Context Protocol server that gives Claude Code, Claude Desktop and other clients one shared SQLite memory, with FTS5 full-text search, filters on kind, domain, project and tags, and version history kept by database triggers. Entry types are registered at runtime and generate their own SQL views, indexes and typed MCP tools with JSON schemas. In daily use as my own knowledge store and CRM, currently 265 entries across 8 types.
github.com/lukaskurz/brainchair
Vineyard row detection from satellite imagery
BSc thesis, Johannes Kepler University Linz, completed 2025, grade sehr gut (1).
I annotated the dataset myself, 1,422 rows across 79 Austrian sites in QGIS, from 30 cm basemap.at orthophotos and INVEKOS parcel boundaries. A U-Net trained on four-channel patches (RGB plus parcel mask) reached 0.99 ± 0.12 m spatial accuracy and 88% row-count accuracy, inside the tolerance agricultural GPS needs. I compared loss functions for thin linear features and defined a row-count metric, because overlap metrics punish one-pixel offsets that do not matter in the field. The best models used centerline Dice, a topology-preserving loss, over the standard Dice that earlier experiments had favoured.
Read the thesis (PDF)
Skills
Python, PyTorch, ONNX, Kotlin, Go, TypeScript; neural surrogates, time-series models, CNNs, on-device inference, LLM agent and RAG workflows.