SYS.CV // NODE: BERLIN_DE (52.5200°N, 13.4050°E) REV 2025.05 · UTF-8

MIRA KOVAČ

[ID: 0x4D4B]

Senior Machine Learning Engineer // Berlin, Germany

7+ years designing low-latency neural retrieval, ranking, and distributed training systems in production. Focused on sub-20ms inference at scale, hardware-aware model quantization, and counterfactual evaluation.

01 //

EXPERIENCE

2017 → PRESENT
2022 — NOW

Senior Machine Learning Engineer @ Zalando SE

BERLIN, DE
  • Architected a two-tower neural retrieval system serving 50M+ active users across 25 markets, replacing lexical BM25 baselines and lifting session GMV by +3.4% (p < 0.001) at 14ms p99 latency.
  • Led INT8/FP16 post-training quantization and custom TensorRT/Triton execution pipelines for transformer re-rankers, cutting GPU fleet provisioning costs by 41% across 120× A10G nodes with zero NDCG@10 degradation.
  • Built an automated counterfactual off-policy evaluation framework (Doubly Robust & clipped IPS estimators) over logged impression streams, shrinking candidate ranking validation cycles from 3 weeks to 48 hours.
2019 — 2022

Machine Learning Engineer @ Delivery Hero

BERLIN, DE
  • Designed and deployed a spatio-temporal Graph Neural Network (PyTorch Geometric) for real-time courier dispatch and ETA prediction across 400+ cities, reducing mean absolute delivery error by 11.8% during peak dinner load.
  • Engineered a low-latency streaming feature pipeline on Apache Flink and Redis Cluster ingesting 95,000 events/sec to compute sub-second courier telemetry features with strict train-serve parity.
  • Implemented continuous feature-drift telemetry using Population Stability Index (PSI) and automated shadow-deployment gates on Kubernetes, eliminating silent regressions during seasonal demand spikes.
2017 — 2019

Machine Learning Engineer @ SoundCloud

BERLIN, DE
  • Trained collaborative metric learning and log-mel spectrogram CNN models over 200M+ tracks to solve cold-start audio discovery, increasing 7-day listener retention on newly uploaded tracks by +6.2%.
  • Migrated offline batch recommendation workflows from legacy Spark MLlib jobs to distributed PyTorch DDP and Faiss (IVF-PQ) vector indices, compressing nightly embedding index builds from 9.5 hours to 38 minutes.
02 //

CAPABILITIES

[CLICK CHIP TO TRACE USAGE]
03 //

SIDE_PROJECTS

OPEN SOURCE
kv-quant

Zero-dependency C++20/CUDA library for grouped 4-bit asymmetric KV-cache quantization in autoregressive transformer decoding.

CUDA / C++20
hnsw-mmap

Memory-mapped Rust implementation of Hierarchical Navigable Small World graphs supporting zero-copy concurrent reads for edge inference.

Rust / PyO3
ips-bench

Reproducible synthetic benchmark suite for evaluating variance-reduction techniques in off-policy contextual bandit estimators.

Python
04 //

PUBLICATION

PEER-REVIEWED

"Debiasing Position and Trust Cascades in Two-Stage Neural Retrieval via Doubly Robust Counterfactual Risk Minimization"

Kovač, M., Lindqvist, H., & Vance, S. · Proc. 17th ACM Conference on Recommender Systems (RecSys '23), pp. 412–423. DOI: 10.1145/3604915.3608819
05 //

TRANSMISSION

ENCRYPTED / DIRECT
GH:github.com/mirakovac GS:scholar/mkovac IN:in/mirakovac