Lead Data Scientist/Machine Learning Engineer in ATB-market

We invite you to join our team Lead Data Scientist/Machine Learning Engineer

Responsibilities:lead the ML direction and roadmap; ensure the transition from ideas to stable production and scaling; demonstrate measurable impact on P&L.

Areas of responsibility: 

• ML initiative priorities, architectural principles, metrics/validation standard; 
• Operational outline: dataset/feature standards, SRM control, roll-out/rollback policies; 
• Reliability/observability of services (SLO/SLI, alerting, cost-aware inference), including edge scenarios; 
• Technical leadership: hiring/mentoring, review, research and knowledge culture; 
• Data governance: PII, access, lineage, model maps/documentation;
• Experimental platform: events, stratification, incrementality;
• Pricing/promo models with elasticities, cannibalization, shelf/stock limits; mission personalization;
• Synchronization of ML goals with budget/plan, transparent impact reporting;

Expected results (OKR examples): 

• NDCG@K > 2 v.p. in personalization (A/B, statistical significance) 
• Maintenance growth due to a relevant promo mix within control corridors.

Requirements:

• 5+ years in ML/DS, 2+ years as a lead/tech lead;
• Proven experience in building and bringing to production on-prem ML services with store/region replication; 
• Cases in CV/Recsys/TS with a proven business effect; production-Python/SQL; MLOps practices, testing, monitoring; 
• Events (streaming) and batch, model monitoring (drifts/stability/degradations), design of experiments, business communications.

Will be a plus: 

• Multimodal features/LLM-signals for cold-start; inventory-aware recsys; promotion optimization; 
• Feature contracts/lineage/metadata management; inference cost optimization (ONNX/TensorRT/quantization);
• Edge inference in hall/SCO, anti-fraud.

Technical stack (on-prem): 

Roles/models

1. CV: Python, PyTorch, OpenCV, Albumentations, YOLOv8–v10 or Detectron2, TrOCR or Tesseract.
2. Recsys: NVIDIA Merlin/Transformers4Rec, implicit (ALS), LightFM, TS reordering/Forecast: LightGBM, CatBoost, XGBoost, N-BEATS, N-HiTS, TFT
3. Vectors: FAISS, pgvector, Milvus | Qdrant. Pandas/Polars for local processing.

NLP/LLM platform 

1. NLP core: Hugging Face Transformers, Datasets, Tokenizers, SentencePiece, spaCy|Stanza (uk), Sacremoses.  
2. LLM serving: vLLM|Hugging Face TGI; TensorRT-LLM|llama.cpp/gguf (by resource profile).  
3. RAG: OpenSearch (BM25) plus re-ranker, chunking and ingest, hybrid search with FAISS|pgvector or Milvus|Qdrant.  
4. Evaluation: ROUGE, BLEU, METEOR, BERTScore, MTEB, Recall@K, MRR, NDCG. Security/PII: Microsoft Presidio. 

MLOps/Serving/experiments

1. MLflow (Tracking/Registry/Serving)
2. Serving: NVIDIA Triton | KServe | Seldon Core | Ray Serve
3. Feature Store: Feast (self-host)

Data and Processing Platform

1. Streaming: Kafka | Redpanda
2. Compute: Spark| Flink
3. Orchestration: Airflow | Dagster
4. Transformations: dbt Core
5. SQL/Storefronts: PostgreSQL, ClickHouse

Data Storage and Architecture

1. Lakehouse: Apache Iceberg | Delta Lake
2. Formats: Parquet, ORC
3. Object storage: MinIO | CEPH

Observability/quality

1. Service observability: Prometheus, Grafana, Loki
2. ML observability: Evidently, whylogs
3. Lineage/directory: OpenLineage, OpenMetadata or DataHub

Infrastructure and security

1. Containerization/cluster: Docker, Kubernetes | OpenShift
2. Security: policy-as-code, Vault Secret Management | Sealed Secrets

Contact person: Kateryna, tel.0984567857 (t.me/KaterynaB_HR)