mira@berlin:~$ whoami

Mira Kovač

Senior Machine Learning Engineer/Berlin, DE/open to staff roles

Eight years of models that ship and stay shipped: ranking on Spotify's home surface, fraud scoring at N26, visual search at Zalando. I like the unglamorous half of the job — feature parity, calibration under label delay, latency budgets, and experiments that survive contact with a real control group.

email mira@kovac.dev github github.com/mirakovac linkedin in/mirakovac tz CET · remote-friendly
01

skills

  • PyTorch
  • Ranking & retrieval
  • Real-time serving — Go, gRPC
  • Distributed training — Ray
  • MLOps — Kubernetes, MLflow
  • CUDA, ONNX, Triton
  • SQL, dbt, Airflow
  • Experiment design & A/B
02

experience

  1. 2022→ present

    Senior Machine Learning Engineer

    Spotify · Berlin · Home & Discovery ranking

    • Replaced the nightly batch ranker on Home with two-tower retrieval plus a gradient-boosted rerank on streaming features: staleness 24h → 12min, +4.1% 7-day artist follows across a three-arm A/B.
    • Wrote the training/serving feature-parity contract that has caught 38 schema-drift bugs before deploy — ranking regressions on Home halved over two quarters.
    • Cut personalization API p99 from 45ms → 26ms by distilling the reranker into a quantized ONNX graph batched on Triton; mentor four engineers on the team.
  2. 2019→ 2022

    Machine Learning Engineer

    N26 · Berlin · Fraud & Risk

    • Owned the real-time card-fraud scorer end to end: 200+ Flink-computed features, p99 under 40ms at ~1.4M transactions/day across six markets.
    • Moved fraud detection off static thresholds to cost-sensitive weekly retraining on delayed label windows — false declines −31% at flat recall.
    • Built the shadow-eval harness and PSI drift alerts that give every model change a week of offline evidence before it touches a live control group.
  3. 2017→ 2019

    Machine Learning Engineer

    Zalando · Berlin · Search & Visual Relevance

    • Fine-tuned a ResNet-50 for garment attribute extraction over 40M product images (neckline, silhouette, sole): mAP 0.71 → 0.84, feeding search facets and the style filter.
    • Trained a LambdaMART reranker on click logs that cut the zero-result rate from 6.2% → 4.8% across DE/AT/CH.
    • Moved nightly training onto Airflow + Kubernetes — 4h → 51min wall time, and it stopped waking anyone up.
03

side projects

  • ckptdiff

    Prints weight drift and sparsity between two PyTorch checkpoints as an ASCII heatmap — no tracking server required.

  • lagwatch

    A 60-line CLI that pages you the moment feature-pipeline Kafka lag crosses your SLO.

  • flatline

    Braille-drawn terminal charts that replay any training CSV, so you can watch curves from inside tmux.

04

publication

2023

“Calibrating Under Delay: Label Latency in Real-Time Fraud Detection”

Kovač, M., Hofmann, L., Petrov, A. — Workshop on Deployable ML Systems, NeurIPS 2023