00 / PROFILE
CURRICULUM VITAE — 2026
MIRA KOVAČ
SENIOR MACHINE LEARNING ENGINEER
- BASE
- Berlin, Germany
- CURRENT
- N26 / Real-time ML
- STATUS
- Open to conversations
Nine years building and operating ML systems across mobility, commerce, and fintech.
I care about clean features, honest evaluation, and models that stay reliable after launch.
VIEW EXPERIENCE ↓
52.5200° N
13.4050° E
[ ML / PROD ]
01 / WORK
Experience
2017 — PRESENT
2023NOW
Senior Machine Learning Engineer
N26
BERLIN, DE
- Lead real-time fraud ranking for card payments across 24 European markets; sequence features and calibrated gradient boosting improved recall by 14% at the same manual-review rate.
- Built streaming features with Kafka, Flink, and Feast, reducing feature staleness from hours to under 30 seconds without training/serving skew.
- Introduced shadow scoring, drift monitors, and signed model releases, giving Risk Operations faster investigations and a one-click rollback path.
20202023
Machine Learning Engineer
Zalando
BERLIN, DE
- Shipped learning-to-rank models for fashion search in TensorFlow, combining query intent and customer context; validated gains through interleaving and online A/B tests.
- Built a two-tower retrieval service for 2M+ fashion articles, using quantized embeddings to hold p95 candidate latency below 35 ms.
- Created reusable evaluation and feature pipelines adopted by three product teams, cutting new ranking experiments from weeks to days.
20172020
Data Scientist, Machine Learning
HERE Technologies
BERLIN, DE
- Developed gradient-boosted travel-time models on anonymized probe data, reducing ETA error by 11% on high-volume urban corridors.
- Rebuilt geospatial feature jobs in Spark and Python, halving daily processing time and making backfills reproducible across regions.
02 / STACK
Skills
TOOLS I SHIP WITH
Python
PyTorch
TensorFlow
XGBoost
Spark
Kafka
Kubernetes
MLflow
03 / AFTER HOURS
Side projects
SMALL, USEFUL, OPEN
P.01
driftline
Opinionated drift reports for tabular models, with PSI, slice-level comparisons, and CI-friendly thresholds.
PYTHON / GREAT EXPECTATIONS
P.02
pocket-signal
A tiny wake-word and command classifier that runs entirely on-device, exported from PyTorch to Core ML.
PYTORCH / CORE ML
P.03
recsys-replay
A reproducible offline playground for comparing retrieval and ranking models against explicit latency budgets.
PYTHON / FAISS
04 / WRITING
Publication
SELECTED PAPER
2022
RECSYS ’22
Beyond Offline Gain: Guardrail Metrics for Learning-to-Rank in Production
Mira Kovač, Jonas Richter
REVEAL Workshop · 16th ACM Conference on Recommender Systems · Seattle, WA
[01]