Data Scientist and AI Engineer based in Vienna. I build LLM pipelines and data analysis tools, mostly around questions where no clean dataset exists and one has to be assembled from open sources first.
MSc Data Science and BSc Chemical and Process Engineering, both TU Wien. English and German (C1), Hungarian (native).
Global Fertilizer Flow —
No public source records which tonne of urea moved from which port to which
port. This approximates the picture from five open datasets: a Sankey tracing
material from mined mineral through intermediates into 23 finished product
categories, sea-route maps coloured by chokepoint, and ~240 ammonia plants —
plus an explicit account of what the approximation does not capture.
Plotly Leaflet searoute FAOSTAT USGS UN Comtrade
Swiss Pharma Landscape —
Market data, ClinicalTrials.gov pipelines and news sentiment for Switzerland's
five largest listed pharma/medtech firms, merged into one self-contained HTML
dashboard. Company mentions are disambiguated from lookalike names, and a
failed fetch reports No Data rather than a fabricated number.
Plotly yfinance ClinicalTrials.gov API News API
Industrial Development Timelines —
Countries industrialize along similar trajectories at different points in
history. Anchoring each country's timeline at the year it crossed a shared
milestone ($2,000 GDP per capita, 50% electricity access) puts them on one
axis — China's path from 2006 read against Japan's from 1970, and on a log
scale fifteen countries collapse into a single band.
Dash Plotly World Bank API Our World in Data
Renewal Radar (CH) —
Upload a Swiss contract PDF, get the notice deadline and contract end as
calendar events. A local LLM with a German rule-based fallback does the
extraction, every field is labelled with its provenance, and nothing leaves
the user's machine.
Docker local LLM PDF parsing iCalendar
Docs-AI Minimal —
Retrieval-augmented Q&A over PDFs behind a REST endpoint: pgvector cosine
search, Redis caching, LangGraph workflows, Docker Compose for local dev.
FastAPI Gemini pgvector Redis LangGraph
OIE LLM Framework —
Two-stage LLM pipeline extracting structured data from German legal text at
80% accuracy, benchmarking GPT-4, LLaMA and DeepSeek against baseline methods.
Python GPT-4 LLaMA DeepSeek
Fine-tuning LLMs for OIE —
LoRA fine-tuning raised LLaMA-3.1's CaRB F1 from 28.1% to 43.0%, a gain of
15 percentage points; GPT-4o reaches 55.0% on the same benchmark. Complete
toolkit for evaluating and fine-tuning LLMs for Open Information Extraction.
PyTorch LoRA LLaMA-3.1 NVIDIA A100