Statistics & Data Science
Student at LMU Munich
I build machine learning and AI projects in Python and R, from deep-learning models that predict when jet engines will fail to a study assistant that answers questions from 7,183 research papers and runs on Azure. I put most of the work into how they are evaluated.
Looking for working-student roles and internships in Munich.
Answers questions only from a curated paper database and cites its sources, or says the database has nothing relevant. Papers come from OpenAlex, pass credibility filters and are stored with embeddings in PostgreSQL. Runs on Azure with a FastAPI backend, a LangGraph tutor, an MCP server and scheduled quiz and refresh jobs.
Remaining-useful-life prediction on NASA's C-MAPSS turbofan data. Compares gradient boosting with an LSTM, a 1D-CNN and a Transformer on identical data, validates by engine rather than by row so no engine leaks into both sets, and explains predictions with SHAP.
Compares the sample covariance with four regularised estimators, including Ledoit–Wolf shrinkage, random-matrix eigenvalue clipping and a PCA factor model, on 428 S&P 500 stocks. Tested in a walk-forward backtest with transaction costs and on a second dataset spanning the 2008 crisis; bootstrap intervals and a deflated Sharpe ratio show which differences hold up.
Binomial tree, Black–Scholes and Monte Carlo written in NumPy without any pricing library, and checked against each other and against no-arbitrage identities. Extended to American early exercise, Asian and barrier options, and a delta-hedging simulation with transaction costs.
A simulated checkout experiment with 40,000 visitors and a known ground truth, analysed with a two-proportion z-test, Welch's t-test, a chi-square test and a power analysis, all written from first principles on NumPy and SciPy.
WHO and UN panel data on 193 countries from 2000 to 2015, analysed in R four ways: regression with LASSO, a mixed-effects model for the repeated-measures structure, ridge-penalised classification after diagnosing quasi-complete separation, and hypothesis tests that check their own assumptions.
A Power BI report on pay in data roles from 2022 to 2025. 422 raw job titles are sorted into 10 role categories by a documented rule, the model is a star schema with 7 dimension tables, and every DAX measure was checked against Python before it was written.