Ультимативный ТИР-ЛИСТ навыков ML-инженера - Что нужно знать для работы в ML/DS

Overview

The speaker presents a skill tier list for getting a first solid ML/DS job, using a mid-level machine-learning engineer in recommender systems (RecSys) as the example. The ranking considers each skill’s combined value for learning, passing interviews, and doing the job. Other specialties and companies may prioritize things differently.

The central advice is to learn the fundamentals and apply them in practice rather than trying to master every technology in advance. Some tools are easier to learn during onboarding, and interview-specific revision can wait until closer to the job search.

Highest-Priority Skills: Critical Multipliers

These skills are especially consequential. Lacking them can make landing a good role difficult, while having them can distinguish a candidate.

• ML system design: Understand the end-to-end lifecycle of an ML product, not just how to train a model. This includes:
  • Translating a business goal into an ML problem and selecting suitable metrics.
  • Collecting, cleaning, and analyzing data.
  • Creating features and training models.
  • Validating models and deciding whether ensembles are useful.
  • Running experiments such as A/B tests.
  • Deploying, monitoring, and retraining models.
  • Recognizing cases where ML may not be needed.
• A coherent résumé and professional narrative (“legend”): Be able to explain projects, responsibilities, decisions, and how your work fits into a real team and business process. The speaker emphasizes understanding and being able to discuss your work, rather than relying on a list of tools or paperwork. He explicitly criticizes forged employment documents.
• Communication and other soft skills: Technical knowledge alone is not enough. Candidates need to explain their reasoning clearly, respond constructively, and communicate well in interviews and with colleagues.
• Effective use of AI tools: Use tools such as ChatGPT and coding agents to research unfamiliar topics, get explanations, and accelerate development. Build foundational understanding rather than outsourcing all learning to AI; AI is most useful when you can review and understand its output.

Foundation: Skills Most Engineers Should Know

• Python: Learn core syntax, functions, loops, and classes, then reinforce them through practical work. The speaker advises against beginning with extensive abstract study of advanced Python or object-oriented design.
• Practical mathematics: Learn enough basic linear algebra, calculus, and related concepts to understand ML methods. Avoid spending months on advanced math or manual calculations before applying ML.
• Core Python data libraries: Become familiar with NumPy, pandas, and Matplotlib. Seaborn and Polars are useful additions. Learn to use documentation rather than trying to memorize every function.
• Scikit-learn: A core library for common ML workflows. Learn it through practice and consult its documentation as needed.
• Classic ML models: Learn linear and logistic regression, including loss functions and regularization; decision trees and random forests; and boosting libraries such as CatBoost, XGBoost, and LightGBM. Understand basic concepts such as bagging, boosting, bias, variance, and ensembles.
• Validation and hyperparameters: Understand cross-validation, folds, and how to choose an appropriate data split. Learn what hyperparameters do and how changing them can affect training and metrics.
• Data analysis: Inspect and understand data before selecting or training a model. “Put everything into a model immediately” is a poor default.
• SQL: Learn the basics, especially joins and window functions. The speaker considers SQL widely useful and relatively quick to learn, but does not recommend studying it to data-engineer depth for this target role.
• Everyday development tools: Know basic terminal commands, Git, and environment management with tools such as Conda. Advanced Git workflows can be learned through project work.
• Basic data collection concepts: Understand where data comes from and when additional data or features may be needed, even if data engineers handle much of the collection.
• Basic CI/CD awareness: Understand what CI/CD is and what it is used for. Detailed expertise in a particular platform is not essential before starting work.

Strong Skills: Valuable Advantages

• Metrics: Understand common classification metrics, confusion-matrix terms, and measures such as precision, recall, F1, ROC-AUC, MSE/RMSE, and log loss. Be able to explain how metrics behave when predictions or values change—not merely recite definitions.
• Feature engineering and preprocessing: Clean and transform data, extract useful information, and create features that can improve model quality. The speaker considers these especially valuable because they can produce gains without the infrastructure cost of more elaborate model ensembles.
• Deep learning and PyTorch: Learn the basics of neural networks, activations, optimizers, regularization, and training. Be able to explain or implement a simple network. The speaker recommends prioritizing PyTorch; TensorFlow/Keras and higher-level wrappers are less important to learn first.
• Model interpretation and calibration: Be able to explain why a model makes a prediction and assess whether its outputs are appropriately calibrated. The speaker says this is often neglected and can indicate stronger engineering judgment.
• Transformers: Understand the architecture and its general uses, especially for NLP, computer vision, and newer recommender systems.
• Recommender systems: A key specialization for the example role. Learn the principles and common approaches used to build recommendation systems.
• Reusable ML pipelines: Organize a complete, reproducible project into a coherent pipeline—from data processing and feature generation to model training, configuration, logging, and validation—instead of leaving disconnected code in notebook cells.
• A/B testing: Understand when an A/B test is appropriate, how test and control groups are formed, and how to think about test duration and evaluation. The speaker emphasizes practical experimental judgment, not just formulas.
• Multimodal ML knowledge: For advanced RecSys work, familiarity with NLP, language models, computer vision, and other modalities can help extract information from product text, images, and other data.
• Embeddings and inference optimization: Understand how embeddings can support retrieval and recommendation, and learn basic ways to improve inference efficiency. The speaker describes deeper inference optimization as a valuable specialization.

Situational or Role-Dependent Skills

These may be useful, but the speaker recommends learning them in proportion to the job or specialty.

• PySpark: Useful in some data environments, but less universally required than SQL.
• Airflow, Docker, Jenkins, and other infrastructure tools: Understand their purpose and basic concepts, but avoid spending large amounts of time mastering a specific company’s setup before joining. Infrastructure varies by organization and can often be learned during onboarding.
• System design beyond ML: Know basic ideas about services, databases, caching, load balancing, and application architecture. The speaker says this is more relevant in some high-load or international roles than in many Russian-market ML interviews.
• Time-series models: Know the basic ideas behind RNNs, GRUs, and LSTMs. They remain useful in some tasks, though they are not needed in every role.
• Search, ranking, matching, uplift modeling, and pricing: These are relevant to particular product problems. For RecSys, search and ranking are especially pertinent; uplift and pricing are more specialized.
• Prompt engineering, agents, and retrieval-augmented generation: Know the basic concepts, but the speaker does not consider advanced RAG expertise a default requirement for a standard RecSys role.
• LinkedIn: Useful for job searching, particularly for international remote work; less essential for the Russian job market.
• English: Valuable because it expands access to documentation, content, tools, and international vacancies. However, the speaker says a lack of English should not stop someone from beginning to study ML.
• Clustering and other unsupervised learning: Understand the intuition behind methods such as k-means and DBSCAN, and the kinds of problems they address.
• Recommendation-related reinforcement learning: Multi-armed bandits and related methods can be powerful when relevant, but are not universal requirements.

Niche or Low-Priority Skills

• AutoML: Can be useful in particular settings, but the speaker warns that it may produce cumbersome models that do not reflect business context. He does not recommend prioritizing it for a first role.
• Cloud platforms: Basic familiarity with services such as AWS or Google Cloud may help, but the platform used varies by employer. Learn the relevant environment when needed.
• Kubernetes and experiment-tracking tools: Useful in some teams, but generally not mandatory advance study for this target role.
• Power BI: More relevant to analytics-oriented roles than to a typical ML-engineering role.
• C++: Worth studying when pursuing a specific low-latency or systems-focused position, but a beginner C++ course alone is unlikely to provide a broad advantage for ordinary ML roles.
• R: The speaker considers it a low priority for the roles discussed.

Suggested Learning Approach

1. Start with broadly reusable foundations: Python, basic math, data libraries, classic ML, SQL, Git, and basic validation.
2. Practice as you learn: Apply concepts to real tasks and projects rather than trying to memorize every API or master theory in isolation.
3. Prioritize understanding and judgment: Learn why a method, metric, or validation strategy is appropriate—not just how to call a library function.
4. Build complete, reproducible projects: Organize code into pipelines instead of relying only on exploratory notebooks.
5. Add stronger differentiators: Develop feature engineering, metrics, deep learning, interpretation, experimentation, and ML system design.
6. Prepare for interviews closer to the job search: Review recurring questions and rehearse explanations once the foundations and practical experience are in place.
7. Learn company-specific infrastructure during onboarding: Focus in advance on concepts and tools likely to transfer between employers; expect implementation details to vary.
8. Use AI as an accelerator, not a substitute for learning: Ask for explanations and research help, and use coding agents when you can assess their work.

Speakers and Sources Featured

• Speaker: Vadim Timakin, the video’s presenter.
• Sources and materials referenced: His complete guide to ML/DS directions, practical ML course, roadmap and training materials, onboarding guidance, and experience mentoring students and interviewing or working with ML engineers. No other speakers are featured.