mira@berlin:~$ cat ./cv.txt
Mira Kovač
Senior Machine Learning Engineer
Nine years shipping ranking, forecasting and fraud models — from research notebook to production serving. Mostly Python and PyTorch, occasionally an XGBoost ensemble that just refuses to be beaten.
## experience
-
2023 — present
Senior ML Engineer @ HelloFresh
- Lead demand forecasting across 12 EU markets; gradient-boosted quantile models cut food waste 18% while holding fill rate flat.
- Own the LLM recipe-recommendation pilot: retrieval over 40k recipes, +9% add-to-cart in a 4-week A/B test.
- Mentor four engineers; introduced model cards and an eval harness now used by six teams.
-
2019 — 2023
Senior ML Engineer @ N26
- Shipped real-time card-fraud scoring: XGBoost ensemble at <50 ms p99, fraud losses down 31% in year one.
- Built the risk feature store on Feast; halved training-serving skew incidents across six models.
- Designed drift monitoring (KS/PSI alerts) that cut silent-degradation MTTR from days to hours.
-
2017 — 2019
Machine Learning Engineer @ Zalando
- Trained a two-tower product-embedding model for fashion search; recall@10 up 14% over 8M SKUs.
- Built Spark feature pipelines for seller-onboarding checks, processing ~2M events nightly.
## skills
- Python
- PyTorch
- XGBoost
- Apache Spark
- Kubernetes
- Feast
- SQL
- ONNX Runtime
## side projects
- seq2seq.wasm—a transformer trained in the browser, 40 KB weights, zero dependencies.
- driftctl—CLI that turns a CSV into terminal sparklines and PSI drift reports.
- repro-curl—one-curl reproducible ML environments; 2.1k stars.
## publication
Kovač, M. & Lindqvist, S. (2022). Calibrated pairwise losses for cold-start product search. In Proc. RecSys ’22, ACM.