Skip to content

Featured Projects

A selection of my data science and machine learning projects. Each one goes end to end, from raw data to a model whose performance has been tested on data it never saw.

  • Predicting Physical Performance from Genome-Wide Data

    Genomic quality control, imputation to ~3.2 million variants and polygenic scores for ~1,200 military recruits. Evaluated over 100 train/test splits with variant weights estimated only on training data, giving a realistic, leakage-free estimate of what genetics can predict.

  • Predicting Global Pine Distributions Under Climate Change

    Ensemble species distribution models (random forests, GAMs, GLMs) predicting how climate change will affect the global distribution of pines by 2070, across hundreds of thousands of climate projections. Validated on independent data from regions never used in training, with AUC ≥ 0.98.

  • EUR/USD Forecasting with Central Bank Sentiment

    Tests whether sentiment from 3.6 million tweets about the ECB and the Fed improves exchange rate forecasts, comparing linear and ensemble models with time-series cross-validation against a random-walk baseline.

  • Modeling Signals of Adaptation Across the Human Genome

    A framework that benchmarks elastic net, random forests, XGBoost and deep neural networks on ~180 GB of genomic data, with nested cross-validation and Bayesian optimization. The best network reached a median R² of 0.83 on held-out data and kept ≈ 0.80 in 25 other populations without re-tuning; explainable AI linked adaptation to genes involved in heat production.

  • Email Triage & Grounded Auto-Reply Workflow

    An LLM workflow that classifies every incoming email in real time and drafts replies grounded in an internal knowledge base. Built with FastAPI, Celery, PydanticAI and pgvector, evaluated with an LLM-as-judge and traced with Langfuse, deployed with Docker on Hetzner.