> hello, world_
Mira Kovač.
Senior Machine Learning Engineer
I build ML systems that hold up outside the notebook.
My focus: recommendations, retrieval, and the engineering
that gets good models into production.
01 / experience
9 years of building
2022present
Senior Machine Learning Engineer
current
Zalando / Berlin
- Led a two-stage product retrieval pipeline using dual-encoder embeddings and a learning-to-rank model; improved add-to-cart rate by 4.6% in a six-week A/B test.
- Cut p95 inference latency from 140 ms to 65 ms through model distillation, batched inference, and a shared feature cache.
- Introduced slice-level evaluation and drift alerts for cold-start products, with canary releases and automated rollback gates.
PyTorch · Spark · Kubernetes · MLflow
20192022
Machine Learning Engineer
Delivery Hero / Berlin
- Built gradient-boosted delivery-time models with calibrated prediction intervals; reduced median absolute error by 18% against the production baseline.
- Reworked training data joins to prevent point-in-time leakage and added feature parity checks between offline training and online serving.
- Owned weekly retraining and shadow deployments across six markets, partnering with operations teams to monitor errors during peak demand.
Python · LightGBM · Airflow · BigQuery
20172019
Data Scientist
SoundCloud / Berlin
- Developed implicit-feedback recommendation baselines and temporal evaluation splits for personalized track discovery.
- Designed listening-session metrics and analyzed online experiments, helping the team distinguish short-term plays from repeat listening.
Python · SQL · scikit-learn · Spark
02 / toolkit
tools, not trophies
- Python
- PyTorch
- SQL
- Recommendation systems
- Vector retrieval
- Experiment design
- ML infrastructure
- Kubernetes
03 / side projects
after hours
001driftwatchA lightweight CLI for catching feature drift before it becomes model drift.+
Compares reference and production samples using PSI and Wasserstein distance, then writes a compact HTML report. Includes minimum-sample checks to keep small batches from triggering noisy alerts.
Python · SciPy · Typer002needleA reproducible benchmark for vector search latency, recall, and memory.+
Runs exact and approximate nearest-neighbor search against the same embedding corpus. Reports recall@10 and p95 latency across index sizes, with fixed seeds and versioned configurations.
Python · FAISS · NumPy003small-batchA practical collection of ML experiments that fit on a single GPU.+
Small, documented experiments in embedding distillation and contrastive learning. Every experiment includes a baseline, an ablation, a compute budget, and notes on what failed.
PyTorch · Hydra · Weights & Biases
04 / publication
notes from the field
TECHNICAL REPORT / 2023Beyond offline accuracy: evaluating
retrieval under catalog shiftMira Kovač · Independent research+
An evaluation protocol for recommendation retrieval when products enter and leave the catalog. The report compares random and temporal splits, measures new-item coverage alongside recall@K, and examines when offline gains fail to carry into production.