Physicist-turned-ML engineer specializing in systems where domain structure matters — time-series, anomaly detection, and generative modeling with physical constraints.
I work at the intersection of physics and machine learning, building systems where domain knowledge isn't decorative but structurally load-bearing. My background spans theoretical particle physics, scientific software development, and applied ML across industry-relevant problems.
Previously I led interdisciplinary research teams across Austria, France, and Germany; managed €1.3M+ in competitive funding; and built SMoldeS, a widely adopted Python toolkit for new-physics searches at the LHC. I now focus on uncertainty-aware prediction, physics-constrained generative models, and agentic ML systems.
Two principles shape how I approach ML. First, structural priors: domain knowledge (physics equations, biochemical constraints, cost models) embedded directly into architecture, where it shapes behaviour rather than just regularising it. Second, scientific rigour: proper baselines, calibrated uncertainty, and evaluation that produces an explanation for why a result happened, not just that it did.
I co-organise the Women in AI, Austria Styrian meetups and a regular speaker at applied ML meetups and international workshops.
2D diagnostic landscape combining reconstruction and physics loss to classify 9 anomaly types without labeled anomalies. Physics-informed kNN outperforms standard baseline by 21 pp on detection rate. Decision framework documented in a 10-part public series reaching 50k+ impressions.
Physics-grounded feature engineering — condition-normalized sensors (KMeans, 6 clusters), rolling statistics, monotone RUL constraint — combined with XGBoost + Optuna tuning. Uncertainty via split conformal prediction (90% coverage). Evaluated on cost-based metrics across all 4 CMAPSS datasets.
Standard VAE collapses to 1D on imbalanced protein data; a differentiable PyTorch GMM penalty recovers all three Ramachandran islands. Latent perturbation across 6 sigma levels confirms 100% win rate in phi stability. Dataset: 1,333–3,335 samples from 5 structurally diverse PDB proteins.
Constrained LLM reasoning architecture that grounds every recommendation in structured audio features — prevents hallucination while preserving human-readable explanations. Hybrid system combining Spotify audio features with semantic topic modeling (HDBSCAN). Explainability is a first-class design requirement, not a post-hoc label.
End-to-end technical walkthrough of building a physics-informed LSTM for anomaly classification: architecture choices, the 2D diagnostic loss landscape, and evaluation across 9 anomaly types.
Read on LinkedInWhat the anomaly detection project looks like from a product angle: the decisions made, what the performance improvement actually means operationally, and how physics constraints change the conversation about model trust.
Read on LinkedInBehind the scenes of creating an AI-generated theme song for TEDxGraz: the process, the tools, and what building something creative with ML teaches about its actual capabilities and limits.
Read on LinkedInOpen to collaborations, speaking invitations, and research or industry roles in ML and physics-informed systems.