01experience
-
2022→ now
Senior Machine Learning Engineer
Zalando · Search & Browse Ranking
- Tech lead for catalog browse ranking: replaced a GBDT ranker with two-tower retrieval and a transformer re-ranker, +2.8% CTR and +1.1% revenue per session in a six-week A/B test across six markets.
- Distilled the re-ranker into a 4-layer student served on CPU via ONNX Runtime — p99 latency from 140 ms to 38 ms and roughly €400k a year less GPU spend.
- Built an IPS-weighted counterfactual replay harness on logged impressions; it now gates ranking releases for five teams.
-
2019→ 2022
Machine Learning Engineer
Delivery Hero · Logistics
- Owned delivery-time (ETA) prediction for 11 markets — quantile LightGBM over ~40M orders a month; cut MAE by 17% and late-delivery complaints by 9%.
- Stood up a Feast + Redis feature store with point-in-time joins, taking feature-to-production time from about three weeks to four days.
- Added PSI-based drift monitoring with retraining triggers; it caught a rider-app logging regression before it reached live forecasts.
-
2017→ 2019
Data Scientist
N26 · Financial Crime
- Built the real-time card-fraud scorer — XGBoost on Kafka streaming features, p99 under 50 ms — lifting recall at a fixed 0.1% false-positive rate from 61% to 74%.
- Wrote the model documentation and validation pack for BaFin review; it became the template for every risk model that followed.
02skills
- Python
- PyTorch
- LightGBM / XGBoost
- Learning to rank
- Spark
- Kubernetes
- ONNX Runtime
- Experimentation
03projects
-
tinyrank
Learning-to-rank losses (LambdaRank, ListNet, ApproxNDCG) in ~600 lines of PyTorch, tested against reference NDCG@k.
-
driftwatch
CLI that diffs two Parquet snapshots and flags feature drift (PSI, KS) — drops into any CI pipeline.
-
ubahn-eta
Quantile forecasts of Berlin U-Bahn delays from the VBB GTFS-realtime feed, with a live calibration plot.
04publication
-
2023
Counterfactual Evaluation of Neural Re-rankers under Position Bias at Catalog Scale