mira@berlin:~$ whoami

Mira Kovač

Senior Machine Learning Engineer·Berlin, Germany·9 years in production ML

Experience

  1. 2022 — Present

    Senior Machine Learning Engineer

    N26 — Berlin

    • Lead the real-time fraud detection platform scoring 4M+ transactions/day; cut false positives 38% by moving from gradient-boosted trees to a calibrated ensemble with online feature updates.
    • Built the feature store on streaming infrastructure (Flink + Feast) that reduced model training-to-serving latency from hours to minutes for 12 product teams.
    • Mentor a team of 5 ML engineers; introduced offline/online evaluation parity gates that caught two regressions before release last year.

  2. 2019 — 2022

    Machine Learning Engineer

    Zalando — Berlin

    • Designed the visual search pipeline (CLIP-style embeddings + approximate nearest neighbor) that now powers image-based product discovery across 50M SKUs.
    • Owned the demand forecasting service for 12 markets; reduced stock-outs 22% with a hierarchical probabilistic model retrained nightly.
    • Shipped an A/B testing framework with CUPED variance reduction, adopted by 30+ teams and still the default experimentation tool.

  3. 2017 — 2019

    Machine Learning Engineer

    SoundCloud — Berlin

    • Built the first version of the audio fingerprinting and recommendation ranking models, lifting 7-day retention for new listeners by 9%.
    • Migrated batch training jobs to Kubernetes, cutting training costs 45% and enabling daily model refreshes instead of weekly.

Skills

PyTorch Python Kubernetes Apache Flink Feature Stores Computer Vision NLP / LLMs MLOps

Side Projects

regr — tiny regression testing for ML models

A pytest-style CLI that snapshots model outputs on golden datasets and diffs them across commits. 1.2k GitHub stars.

berlin-cycling — predicting bike lane congestion

Open-data project forecasting hourly cyclist volumes from weather and traffic sensors; live map at berlin-cycling.dev.

flashcards-ml — spaced repetition for ML concepts

A spaced-repetition deck generator that turns any arXiv paper into 20–30 self-test cards using an LLM pipeline.

Publication

"Calibrated Ensembles for Real-Time Fraud Scoring"

Kovač, M., et al. — KDD 2024, Applied Data Science track. Paper · Slides

Contact

mira.kovac@example.com github.com/mirakovac linkedin.com/in/mirakovac Berlin · CET · open to interesting problems