Mira Kovač

machine learning engineer/search relevance

• Rebuilt query understanding for fashion search with a character-level tagger; zero-result queries dropped from 9.4% to 3.1%.
• Shipped a two-tower retrieval model over 40M article embeddings, lifting similar-items click-through by 6.2% in a four-week test.
• Built the offline evaluation harness (NDCG@10, recall@100) that became the release gate for every search model.

senior machine learning engineer/podcast ranking

• Led ranking for podcast recommendations on Home; replaced a GBDT ranker with a sequence transformer and gained +3.8% podcast consumption.
• Cut p99 feature latency from 180 ms to 45 ms by moving feature assembly into a Go service with precomputed embeddings.
• Added editorial and freshness constraints to the ranker, cutting already-consumed impressions by 27% with no engagement loss.

senior machine learning engineer/talent matching

• Own candidate–role matching for 6,000+ SMB customers; hard-negative mining and a tuned encoder raised recall@20 by 19%.
• Introduced a model registry and shadow deploys on Kubernetes, taking rollback time from hours to under ten minutes.
• Mentor three engineers and run the applied-ML review board; the team ships roughly one model a month.