MMOOZZGG.com

LLM

LLM engineer: what the job is and how to learn it for free

7 min read9 sources

An LLM engineer builds working systems on large language models: search over company documents, AI agents, quality checks of answers and tracking of token cost. On 16 October 2026, a search by job title on hh.ru, Russia’s largest job site, found 120 vacancies in Russia with the word LLM. Below are the tasks of the job, how it differs from neighboring roles, the skills and free university courses.

On this page · 6 sections

What an LLM engineer does

An LLM engineer usually does not train a model from scratch. They take a ready model, through an API or on their own server, and build it into a process: connect data, tools and rules, and then check that the answers are good enough for the job.

First typical task: RAG, answers from company documents. Lewis and co-authors described the approach on 22 May 2020: the model answers not only from its own memory but also from chunks that a search engine found in a database. In practice the engineer cuts documents into chunks, computes vectors and sets up the search.

Second task: agents. The model gets tools, such as a query to the CRM or 1C, and decides itself when to call them. The engineer describes the tools, the access rights and the moment when the agent hands the case to a person.

Third task: quality evaluation. A set of test questions with reference answers is run after each prompt edit or model change. Part of the grading goes to another model: in the 2023 work of Zheng and co-authors, GPT-4 as a judge agreed with human ratings in more than 80% of cases, about as often as people agree with each other.

Fourth task: money. Models charge for tokens, the pieces of words a model counts text in. The engineer calculates the cost of one answer, removes extra context and sends simple requests to a low-cost model.

How it differs from an ML engineer and a prompt engineer

An ML engineer trains models on company data: demand forecasts, scoring, finding defects in photos. The work starts with a data sample and ends with a model metric.

A language model engineer more often takes a ready model and builds a system around it: data, tools, checks, monitoring of answers. They bring in training only when there is no other way, for example for fine-tuning on your own examples.

A prompt engineer writes and edits requests to a model. For an LLM engineer this is one skill, not the whole job: beyond the prompt, they are responsible for search, code, tests and costs.

On the market the borders are blurred. On 16 October 2026, the word LLM appeared in the text of 1,265 vacancies on hh.ru across Russia, but in the title of only 120: models are mentioned in vacancies for testers, DevOps engineers and developers too.

Skills: what you need to know

The base: Python and work with model APIs. Requests, streaming answers, tool calls, handling errors and limits come up every day.

How the model is built: tokens, the context window, temperature, the attention mechanism. Without this it is hard to understand why an answer got cut off or became random. How a model works is covered on a separate page of this section.

Search and data: embeddings, that is, vectors of a text’s meaning, vector databases, splitting documents into chunks, keyword search next to search by meaning. This is the core of RAG from the 2020 paper.

Evaluation: test sets, metrics, a model as a judge and manual review of disputed cases. A judge model is trusted only after it is checked against human ratings on your own set of questions.

Security: the agent’s access rights, protection against prompt injection, and 152-FZ, Russia’s personal data law, if the model sees customers’ personal data.

English good enough to read documentation. Papers, model descriptions and library guides come out in English first.

How to enter the profession

Most people come to the job from Python development, data analysis and ML. The first know code and APIs, the second metrics, the third models.

You can build the path in three steps. First, a first agent in one evening on a ready model, then RAG over your own documents, then a test set that catches errors after each edit.

A portfolio weighs more than a certificate. An employer needs a project with a clear metric: what share of questions the system closes, what one answer costs, how often it makes mistakes.

Demand for the skill is growing beyond engineering roles too. According to hh.ru, in the first half of 2026 employers posted more than 21,600 vacancies requiring skills with neural networks.

Where to learn for free

Below are university courses whose pages we opened on 16 October 2026. Programs change every semester, so check the terms on the course page before you start.

Stanford CS336 “Language Modeling from Scratch”, spring 2026: a language model from scratch. Tokenization, the transformer, training and scaling laws, inference, evaluation, fine-tuning and RLHF; lecture recordings are on YouTube.

Stanford CS224N, natural language processing with deep learning. The course covers the attention mechanism, transformers and LLMs, and the full set of videos for the 2024 class is free on YouTube.

MIT 6.S191 “Introduction to Deep Learning”, 2026: an intro course in deep learning with a topic on large language models. The materials are free under the MIT license, and the practice includes a lab on fine-tuning an LLM.

MIPT Deep Learning School, in Russian. You can study for free, and there is a paid Premium subscription; the second part is about text: embeddings, the attention mechanism, transformers and large language models.

AI courses for staff and managers are collected in our training section. Which universities in the world are strongest in AI, and what programs they have, is shown in the ranking of the world’s universities in AI.

Does a company need its own LLM engineer?

You need your own engineer when a language model becomes part of the product or when the company has many processes on models. Then someone must watch quality, costs and model updates every day.

For one or two processes, a staff specialist often costs more than the result. According to hh.ru data as of 17 July 2026, data scientists were offered on average $2,400 a month in vacancies with neural network skills, and programmers $1,500.

There is a third option: an external AI director. They choose the processes, set tasks for the contractor, check the metrics and answer for the result to the CEO, while the agents themselves can be ordered turnkey.

If you already have an engineer on staff, the AI director does not replace them. They give the engineer a queue of tasks, each with a price and a metric for the business.

Questions and answers

How much does an LLM engineer earn?

hh.ru, Russia’s largest job site, publishes no separate statistics for this job, and salaries in vacancies change every week. The closest guide: in vacancies with neural network skills in the first half of 2026, data scientists were offered on average $2,400 a month and programmers $1,500. For fresh numbers, search hh.ru for vacancies with LLM in the title.

Do you need a university degree?

Skills and projects matter more for entry: Python, work with model APIs, RAG and quality tests. University math and statistics help you understand why a model makes mistakes. Free courses from Stanford, MIT and MIPT cover the theory, and a portfolio with a metric shows an employer what you can do.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. hh.ru, vacancy search by title, 16 October 2026 (hh.ru)
  2. hh.ru, vacancy search, 16 October 2026 (hh.ru)
  3. CNews, hh.ru data, 17 July 2026 (cnews.ru)
  4. Stanford University, CS336, 2026 (cs336.stanford.edu)
  5. MIT, Introduction to Deep Learning, 30 March 2026 (introtodeeplearning.com)
  6. Stanford University, CS224N, 2026 (web.stanford.edu)
  7. MIPT Deep Learning School (Phystech School of Applied Mathematics and Informatics), 2026 (dls.samcs.ru)
  8. arXiv, 22 May 2020 (arxiv.org)
  9. arXiv, 9 June 2023 (arxiv.org)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us