ML Engineer: market

Vacancy

an open job that cleared the quality gate and lists at least two technologies

Observed candidate

a candidate whose stack contains this technology; an aggregate — not a single record leaves the perimeter

Matchable candidate

not counted yet

Window

jobs open at the moment the snapshot was built

Junior ML Engineer — rare (the market starts with Data Analyst or Backend → ML). The Junior → Middle jump — after mastering the first production ML pipeline end-to-end (feature engineering → training → deployment → monitoring). Middle → Senior — multi-pipeline ownership + LLM/RAG experience + scaling distributed training + cost optimisation. Senior → Staff/Principal — ML platform architecture, multi-team ML strategy. Career flow: Data Analyst (1-2 years) → Junior ML Engineer (1-2 years) → Middle (2-3 years) → Senior → either Staff ML Engineer (deep technical), ML Tech Lead (management), a move into Research (PhD track), or ML Founder/CTO at a startup.

Top 5: python, cloud, pytorch, llm, aws. Python is the monopoly (R is barely seen in production ML 2026). Deep Learning frameworks: PyTorch dominates 2026 (70%+ of production), TensorFlow legacy + still big in enterprise, JAX for research edge cases. Classical ML: scikit-learn + XGBoost + LightGBM + CatBoost — base, must for tabular data. Data manipulation: pandas + NumPy + Polars (modern, rising 2026 — 10-100× faster), PyArrow for cross-framework data. Computer Vision: OpenCV, torchvision, Detectron2, MMDetection, Ultralytics YOLO. Experiment tracking: MLflow (open-source standard) or Weights & Biases (premium UI) or Neptune. Data version control: DVC. Orchestration: Airflow + Prefect + Dagster (one deeply). Deployment: Docker + Kubernetes mastery, FastAPI + Triton Inference Server + BentoML + TorchServe (one deeply). Cross-framework inference: ONNX. Feature stores: Feast (open-source) or Tecton (managed). Distributed compute: Spark + Ray + Dask. Cloud-managed ML: AWS SageMaker / GCP Vertex AI / Azure ML (one deeply). Senior extras: quantization (PyTorch quantization, bitsandbytes), distillation, LoRA/QLoRA fine-tuning, vLLM/TGI/SGLang for LLM serving, Optuna for hyperparameter tuning.

Data Scientist — focus on business problem framing + statistical modelling + A/B testing + insights communication. Stack: pandas + sklearn + Jupyter + SQL + Tableau/Looker. Less infrastructure-heavy. See Data Scientist. ML Engineer (this page) — focus on production ML pipelines + deployment + scaling. Stack like DS + Docker/K8s + MLflow + Airflow + cloud-managed ML. Research Engineer / Scientist — focus on new models, papers, novel architectures. Often PhD track. Stack: PyTorch deep + JAX + experiment-heavy infrastructure. See Research Engineer / Scientist. MLOps Engineer — focus on ML platform infrastructure (feature stores, model registry, monitoring, observability, cost optimisation). Stack: K8s + Terraform + MLflow/Kubeflow + Prometheus/Grafana + cloud. See MLOps Engineer. Career switch ML Engineer ↔ DS — 1-3 months. ML Engineer ↔ MLOps — 2-4 months (if K8s base exists). Research path requires PhD or 3+ years of deep specialisation.

Reference stack for production ML 2026: 1) Data ingestion — Airflow/Prefect/Dagster pulls raw data from source systems (DB/Kafka/S3). 2) Feature engineering — Spark/Pandas/Polars transforms → Feature Store (Feast or Tecton — centralised storage, point-in-time correctness, online+offline serving). 3) Experimentation — Jupyter + W&B/MLflow tracking + DVC for data version control + Optuna for hyperparameter search. 4) Training — Ray Train or PyTorch Lightning + distributed (multi-GPU/multi-node), checkpointing in S3/GCS, quantization (INT8/FP16) for inference cost. 5) Model registry — MLflow Model Registry or SageMaker Model Registry (versioning + stages: staging/production/archived). 6) Validation — automated tests (data drift, prediction quality), shadow deployment, A/B test setup. 7) Serving — Triton Inference Server (NVIDIA, multi-framework + dynamic batching) or BentoML or TorchServe or vLLM (for LLM). FastAPI wrapper for API contract. 8) Deployment — Kubernetes (production standard) + Helm + ArgoCD for GitOps. 9) Monitoring — Prometheus + Grafana (latency/throughput/error rate) + custom drift detection (Evidently AI/WhyLabs) + cost dashboards. 10) Feedback loop — log predictions + ground truth → retraining trigger. ML platform for the whole organisation — Kubeflow or Metaflow or SageMaker Pipelines. LLM-specific stack: vLLM/TGI/SGLang for serving + prompt engineering (LangChain/LlamaIndex) + RAG (vector DBs Pinecone/Weaviate/Qdrant) + fine-tuning via LoRA/QLoRA + evaluation (RAGAS, DeepEval, custom benchmarks).

Yes, 29% of ML Engineer jobs are full-remote or hybrid. Production ML is cloud-based work (training on cloud GPUs, serving in Kubernetes, data in the warehouse). Outsourcers (EPAM AI Practice / Luxoft AI) — almost always remote on US AI projects. Russian product companies — hybrid or remote after probation, exceptions for security-sensitive ML (Sber AI banking). Banks (Sber AI / VTB) — hybrid/office due to compliance. International AI companies — full-remote standard, premium bands. Relocant hubs: Poland (Warsaw/Krakow — AI-friendly), Germany (Berlin AI cluster), Canada (Toronto — Vector Institute), Israel (Tel Aviv AI cluster), Serbia, Georgia, UAE (Dubai — AI tax-friendly). English for international AI remote — must (premium +30-50%).

ML Engineer (general) — can work in any ML domain (recsys, NLP, CV, fraud detection, ranking). Stack is universal. Applied ML Engineer — focus on a specific product ML task + business-metrics ownership (CTR, conversion, revenue lift). Stack is the same, but adds product thinking + A/B testing mastery + stakeholder communication. Pay comparable (median), but at product companies Applied ML is usually somewhat higher at Senior thanks to business-impact ownership. ML Platform Engineer (a separate sub-niche — closer to MLOps) — builds infrastructure for other ML Engineers (Feature Store, MLflow infrastructure, training platform). Career choice: ML Engineer if universality and portability matter most, Applied ML Engineer if product impact is interesting, ML Platform Engineer if infrastructure and tooling. All three are legitimate Senior+ paths, not better/worse.

Roadmap: 1) Python deep to Backend Middle level (clean code, design patterns, async). Books: "Fluent Python" Ramalho, "Effective Python" Slatkin. 2) Math — linear algebra + calculus + probability/statistics at the level of "understanding what's under the hood of backprop". MIT 18.06 Linear Algebra (Strang), 3Blue1Brown "Essence of Linear Algebra". 3) Classical ML — scikit-learn + XGBoost + LightGBM mastery. Andrew Ng "Machine Learning Specialization" (Coursera). Book "Hands-On Machine Learning" Géron (must-read 2026). 4) Deep Learning — PyTorch deep (forward/backward/optimisers/training loops from scratch). "Deep Learning Specialization" Andrew Ng. "Deep Learning with PyTorch" Stevens. 5) NLP/LLM track: Hugging Face course (free) → fine-tuning small LLMs (Mistral 7B/Llama 3 8B) with LoRA/QLoRA → build a RAG application with LangChain + Qdrant. 6) Computer Vision track (alternative): torchvision + Detectron2 + Ultralytics YOLO + custom dataset training. 7) MLOps basics: MLflow + DVC + Airflow + Docker + Kubernetes for ML serving. "Designing Machine Learning Systems" Chip Huyen — must-read for production ML. 8) Cloud ML: one of AWS SageMaker / GCP Vertex AI / Azure ML — go through one e2e tutorial. 9) Distributed training: Ray Train + PyTorch DDP. 10) Pet project: full e2e ML system (data pipeline + feature store + training + MLflow tracking + Kubernetes deployment + monitoring) on an open dataset (Kaggle). International (EN): fast.ai "Practical Deep Learning", Hugging Face NLP Course, DeepLearning.AI (Andrew Ng). Classic books: "Pattern Recognition and Machine Learning" Bishop, "The Hundred-Page Machine Learning Book" Burkov. Conferences: NeurIPS, ICML, ICLR (papers — must reading for Senior). Data Analyst or Backend Middle → ML Engineer Junior — 8-14 months.

1904 active open ML Engineer positions — the largest ML/AI segment. Geography: Russia / Poland / remote. Sources: hh.ru, Habr Career, getmatch, Djinni, LinkedIn (huge international AI segment), NoFluffJobs/JustJoin.it (Poland — AI-friendly), Telegram, career pages of EPAM AI / Luxoft AI / Andersen AI, specialised boards (aijobs.net, ai-jobs.net, builtin.com/jobs/ai, jobs.lever.co for AI startups), Y Combinator Work at a Startup. The real market is broader thanks to a huge international remote segment. Time to close a Senior ML Engineer role — 6-12 weeks (longer than Backend thanks to rare-skill combination).

A Senior ML Engineer owns the full production ML cycle + technical leadership. Programming mastery: Python Backend Senior level (clean code, async, profiling, memory optimisation). Math foundations: linear algebra + calculus + probability at the level of "reading papers without blocks". Classical ML mastery: deep feature engineering, model selection rationale, calibration, evaluation metrics trade-offs, business-aligned model design. Deep Learning mastery: PyTorch deep (custom Layers/Modules/Optimizers), distributed training (DDP, FSDP), mixed-precision, quantization (INT8/FP16), distillation, ONNX export. LLM mastery (2026 must for Senior): fine-tuning via LoRA/QLoRA, prompt engineering, RAG architecture (vector DBs Pinecone/Weaviate/Qdrant), evaluation (RAGAS, DeepEval), vLLM/TGI/SGLang for serving. MLOps: MLflow / W&B mastery, DVC, Airflow / Prefect / Dagster, Feature Stores (Feast / Tecton), Docker + Kubernetes mastery for ML serving, Helm + ArgoCD. Cloud ML: one of AWS SageMaker / GCP Vertex AI / Azure ML deeply. Distributed compute: Ray + Spark mastery. Monitoring: Prometheus + Grafana + Evidently AI (drift detection) integration. Cost optimisation: GPU utilisation, batch sizing, model quantization, caching strategies. System design for ML: design ML systems on whiteboard at 100M+ user scale (recsys / fraud / ranking / chatbot). Soft: papers reading (NeurIPS / ICML / ICLR), technical writing (ADRs, design docs), code review for multiple teams, mentoring Middle ML Engineers, communication with product/data/backend teams, A/B test design + statistical analysis. English for Senior+ MUST — the ML/AI community and papers are in English. Bonus: open-source contributions (PyTorch / Hugging Face / LangChain) — sharply increase market value.