226 job postings reveal: 75% of AI specialists are expected to work in the office
An analysis of 226 job postings on hh.ru reveals not one unified “AI market,” but three distinct professions: model development, building LLM applications, and maintaining ML infrastructure. Of 144 relevant postings, 104 concern models, 34 applied AI, and only six MLOps. The key takeaway for candidates is simple: look at the responsibilities, not the fashionable job title.
AI-processed from Habr AI; edited by Hamidun News
Today, the title “ML Engineer” can conceal three almost entirely different professions. One vacancy may require training models and calculating metrics; another, building a RAG service on top of a ready-made LLM; and a third, distributing computation across GPUs and maintaining infrastructure. A list of fashionable technologies therefore tells candidates little by itself.
Researchers analyzed 226 hh.ru postings saved on November 6, 2026. After classification, 144 vacancies were deemed relevant: 104 involved ML and DL model development, 34 applied AI systems, and another six MLOps. This is not a census of the entire market, but a snapshot of one sample on a specific day. Even so, it clearly shows where job titles diverge from the actual work.
Three markets under one banner
The first track is model development. It includes data preparation, training, quality evaluation, and experimentation. Computer vision, traditional Data Science, scoring, recommendation systems, and NLP are prominent in the sample. This is the largest category, accounting for 104 of the 144 relevant postings.
The second track is applied AI. Employers here often do not need someone to train a proprietary language model. They need an engineer who can connect a ready-made LLM, build document search, and add RAG, agents, APIs, logging, and answer evaluation. There are 34 such vacancies. This work is closer to product development: value comes not from the model alone, but from a reliable workflow built around it.
The third track is MLOps and ML infrastructure. Infrastructure was the primary responsibility in only six postings. They covered GPUs, model deployment, observability, and pipeline security. The figure is small, but it should not be read as a lack of demand: infrastructure responsibilities also appear within model-developer vacancies.
A simple example: two postings may both be titled “AI Engineer.” In the first, the candidate fine-tunes a model and compares metrics. In the second, they call a third-party model through an API, retrieve documents for context, and handle failures. The title is the same, but the portfolio, interview, and day-to-day work will differ.
What employers put in their requirements
Among the 144 relevant vacancies, Python is mentioned in 55, or 38.2%. LLM appears in 54 postings, or 37.5%; AI agents in 35, or 24.3%; PyTorch in 26, or 18.1%; and RAG in 21, or 14.6%. These percentages should not be turned into a ranking of mandatory skills: the researchers had access only to titles and saved excerpts from descriptions. If a technology did not appear in an excerpt, that does not mean the employer does not use it.
The market leaves newcomers a narrow entry point. Only six relevant vacancies are formally marked “no experience required.” Another 61 require one to three years of experience, 64 require three to six years, and 13 require more than six years. At the same time, “no experience” does not mean “no knowledge”: even a junior specialist needs the fundamentals, working tools, and a project that can be discussed in terms of decisions, metrics, and constraints.
Salary data is even scarcer. Compensation is published in only 32 of the 144 relevant postings. The sample mixes rubles, tenge, and dollars; pre-tax and after-tax amounts; minimum figures and complete ranges. It therefore cannot produce an honest average salary. Individual offers are useful as reference points, but not as a market benchmark.
Practical takeaway
Candidates should first choose the type of work and only then assemble their stack. Model development requires data, metrics, validation, and reproducible experiments. Applied AI requires integrations, error handling, answer evaluation, source citations, and monitoring. MLOps requires deployment, observability, compute infrastructure, and security.
A portfolio should likewise be built around more than the phrase “built a chatbot.” A completed system is more compelling: a clear task, a data source, a baseline for comparison, a chosen metric, constraints, and a launch method. One more sobering signal: 75% of relevant vacancies expect on-site work, while remote work is specified in only 22.9%. The AI job market proved not only demanding in terms of experience, but also significantly less remote than industry conversations might suggest.
- Oct 13, 2026Remote Jobs for IT Analysts Are Shrinking: Share Falls From 33% to 26%
- Oct 12, 2026OpenAI Expands Its Dublin Headquarters: New Jobs at the EU Office
- Nov 6, 2026Baidu Open-Sources ERNIE 5.5 Flash Weights: 14 Billion of 280 Billion Parameters Are Active per Token (5%)
- Jul 12, 2026Paris voice AI startup Gradium raises $100M with Nvidia backing
Key Questions Answered
What are the key details on Three markets under one banner?
The first track is model development. It includes data preparation, training, quality evaluation, and experimentation. Computer vision, traditional Data Science, scoring, recommendation systems, and NLP are prominent in the sample. This is the largest category, accounting for 104 of the 144 relevant postings. The second track is applied AI. Employers here often do not need someone to train a proprietary language model.
What employers put in their requirements?
Among the 144 relevant vacancies, Python is mentioned in 55, or 38.2%. LLM appears in 54 postings, or 37.5%; AI agents in 35, or 24.3%; PyTorch in 26, or 18.1%; and RAG in 21, or 14.6%. These percentages should not be turned into a ranking of mandatory skills: the researchers had access only to titles and saved excerpts from descriptions. If a technology did not appear in an excerpt, that does not mean the employer does not use it. The market leaves newcomers a narrow entry point.
What are the key details on Practical takeaway?
Candidates should first choose the type of work and only then assemble their stack. Model development requires data, metrics, validation, and reproducible experiments. Applied AI requires integrations, error handling, answer evaluation, source citations, and monitoring. MLOps requires deployment, observability, compute infrastructure, and security.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.