Experience
-
2023 — PRESENT
Helsing
Senior Machine Learning Engineer · Berlin
- Rebuilt the credit-risk scoring model from a single logistic regression to a gradient-boosted ensemble, cutting annualized loss by 1.4 pts while holding calibration within 1%.
- Shipped a near-real-time feature store (Kafka + Redis) feeding 40+ features into the risk pipeline, reducing decision latency from 400 ms to 60 ms.
- Owns the end-to-end MLOps pipeline — training, evaluation, canary deployment — for the six models in production.
-
2020 — 2023
Zalando
Machine Learning Engineer · Berlin
- Built a two-tower retrieval + fine-ranking stack for on-site search, improving click-through rate by 6% over the previous linear baseline.
- Deployed a near-real-time CTR prediction pipeline handling 2M requests/day with p95 latency under 50 ms.
- Cut inference cost by 30% via INT8 quantization and ONNX export of the ranking model.
-
2017 — 2020
Deutsche Telekom
Data Scientist · Munich
- Developed churn-prediction models for mobile subscribers, reducing churn by 3% among the top decile of the risk score.
- Built a classification pipeline routing 1.2M support tickets/month into 40 intent categories, with 92% of tickets reaching the right team.
- Wrote the team's A/B testing playbook and ran 30+ experiments on pricing and retention.
Skills
- Python
- PyTorch
- TensorFlow
- SQL
- Kafka
- Kubernetes
- MLflow
Side Projects
- gridwalk — a Rust RL agent for 2D-grid pathfinding, built to compare PPO against hand-rolled value iteration.
- papertrail — a self-hosted arXiv + BibTeX tracker with a terminal UI: one key to search, one to cite.
- noise-floor — a time-series anomaly detector for home IoT sensors, trained on 18 months of readings.
Publication
Kovač, M. & Richter, S.
"When Trees Beat Networks: A Systematic Comparison of Gradient-Boosted and Neural Models for Tabular Credit-Risk Data."
Journal of Machine Learning Research, 25(12), 2024.