How many stages there will be
The number of technical sections is set by the level the candidate is applying for, not by the company. In hiring, that level is called the grade. The higher it is, the more sections there are and the more weight architecture and experience carry in them. Some sections are not offered to junior candidates at all, so the set of sections shows which level a candidate is being considered for.
Yandex described this in October 2025 when it announced its updated hiring process: junior and middle developers go through 2–3 technical sections, seniors through 4–5. The architecture section is held only with those who have commercial experience designing high-load distributed systems. Avito puts the same logic more directly: the design section is not held with everyone, and its playbook calls it “the main differentiator for reaching level E5” — one of the steps in the company’s internal grade ladder.
Recruiter screening
At the first stage, the candidate is assessed on formal criteria: the salary range, the work format, how well they match the job requirements. Nobody tests skills at this point, yet this is where the largest share of candidates in the whole funnel drops out. The reason is volume. A recruiter at the company Infostart published her annual funnel: 43,000 applications for 171 openings — about 250 people per position on average, and up to a thousand for popular roles. A flow like that cannot be sorted by hand, so companies hand part of the selection over to software.
Automatic screening of applications
Some applications are filtered out before a person sees them. In Russian hiring this is not a retelling of Western horror stories about robot recruiters: the mechanism is built into the job board itself, takes a minute to switch on and is free for the employer.
Two different mechanisms are often confused. Ranking only changes the order of results and excludes no one: hh.ru’s “smart search” weighs more than 700 résumé parameters — experience, skills, salary expectations, location — and moves suitable candidates up. Huntflow AI works the same way: it recommends candidates from the database and leaves the decision to the recruiter. Filtering works differently — a rule fires on its own and ends in a rejection.
On hh.ru, filtering is handled by automatic sorting of applications (in Russian). The documentation says every search filter except text fields is available for it, and explicitly lists education, language skills, location and willingness to relocate, work experience and desired income. The system itself marks an unsuitable candidate as “Reject” and sends a template email. The same page gives the timings by which the mechanism can be recognised from the outside: a rejection arrives 15 minutes after the résumé is sent, an invitation within one to two. A rejection that comes within the same hour almost certainly means nobody read the résumé.
The set of filters does not end there. Besides job boards, applications are processed by hiring software that a company installs in-house — known in the industry as an ATS. Potok describes its automatic screening right on its product page (in Russian): the robot “will check all résumés and filter them by age, sex and citizenship”, reject those who don’t match and send them a “polite rejection” from a template.
This leads to a conclusion opposite to the common one. The advice to “make your résumé readable for the robot: no tables, columns or images” comes from Western articles, where résumés are sent as a file and parsed by software. It hardly applies to an application through hh.ru: the employer sees the filled-in fields on the platform, and an attached file plays no part in this scenario at all. Layout starts to matter only where the résumé is sent as a file — by email, in a messenger or through a company’s careers portal.
Attitudes to automatic screening in Russia are consistently negative: according to a SuperJob poll of 1,600 respondents conducted in May 2026, 57% view it negatively, 50% believe using AI to screen résumés should be banned by law, and 8% view it positively.
The recruiter call
Applications that pass the automatic filter get a few seconds of a person’s attention. By the same recruiter’s estimate, the first look at one résumé takes 5–7 seconds. This look filters out another 60–70% of applications. Reading each one carefully is impossible, so the recruiter decides on what is visible at once.
Candidates are lost at this stage for another reason as well. A company often works from its own candidate profile, created at an earlier contact, and the data in it goes stale. A candidate who described his interviews in February 2026 ran into this at Yandex: the recruiter negotiated from a salary figure in the old profile — one the candidate had not given this time. The conversation about money started from the wrong number before the first technical section. Screening and the HR interview are covered separately.
Take-home assignment
The take-home assignment is disappearing from hiring for experienced developers for two reasons, and neural networks are only the second. The first is that developers stopped doing it. According to a SuperJob poll of 1,400 respondents conducted in July 2025, 27% of programmers never apply to vacancies that include a take-home assignment. Over four years that share grew by 7 percentage points. The aversion is specific to development: among testers, 10% answer the same way.
Neural networks finished the job. An assignment that can be completed with one chatbot prompt no longer says anything about its author’s skill, and companies are moving to live formats where the candidate’s reasoning is visible. In the published descriptions of the process at Yandex and Avito there is no take-home assignment at all.
The format survives where selection has to be done at scale: in internships and in hiring entry-level developers, where thousands of applications come in and manual review is impossible. How many hours is considered normal, when a take-home is paid and what a proper refusal looks like — see the take-home assignment breakdown.
Language and platform section
The topics in this section are predictable and hardly differ between companies: language primitives, framework features, concurrency, memory models. The section takes an hour to an hour and a half.
What differs is the conditions the candidate works under, and these determine which skill is actually being tested. When the code cannot be run, the interviewer is testing the ability to find bugs by reading: the usual “write some more and run the tests” loop does not work in such an environment, and debugging has to happen in the head. When the environment is a working one, a mistake costs less, but more is expected of the result. Each company has its own rules and publishes them in advance.
| Company | Environment | What is allowed |
|---|---|---|
| Yandex, C++ and Go | The company’s online editor, with no compiler or autocomplete. The candidate’s own IDE is not allowed. | Internet search is allowed. AI assistants are not. |
| Yandex, Java and Python | The candidate’s IDE with screen sharing. | Documentation, the standard library and search are allowed. AI assistants (Copilot, ChatGPT, Cursor) are banned by name, and the company publishes instructions for disabling them in the IDE. |
| Yandex, frontend | Online editor; code runs and is debugged in the browser developer tools. | Searching the internet for the task and getting outside help are not allowed. |
| T-Bank | The company’s online editor. According to candidate reports, SQL executes and the result is visible. | The general rule is “try not to peek”: about search engines and neural networks, the company warns that “it shows”. Documentation is explicitly allowed only on the Scala track. |
| Avito | Any language suited to the position, or pseudocode. | Library functions are not allowed. |
| VK | The candidate’s IDE with screen sharing; a working environment where code runs and is tested. | No public rules on AI or search. |
| Sber | Live coding for 15–30 minutes, held with the candidate’s consent; for some roles it is not held at all. | No public rules on AI or search. |
The restrictions do not add up to a single logic even within one company: at Yandex, as the table shows, the rules of the backend and frontend tracks contradict each other. The conditions are set by the specific track, so the experience of a colleague who took the section at the same company says nothing about them. Live coding formats are covered in a separate article.
Algorithms section
This is the only stage for which companies publish assessment criteria, and what they describe is the solving process. They look at how quickly the candidate finds an algorithm, how clearly they explain it, whether they can work out its complexity, how many bugs they make in the code and how they fix them. A missed edge case is not a failure in itself. The interviewer in this section stays in the background and steps in only as a last resort, so the candidate has to reason out loud on their own: if they stay silent, there is nothing to assess.
The format is stable: two tasks, an hour to an hour and a half, a plain editor without autocomplete. Yandex, which renamed the section “Basic technical skills check”, has even published a list of LeetCode tags it recommends preparing with — Array, String, Tree, Binary Search, Hash Table, DFS, BFS, Two Pointers, Stack, Backtracking, at easy and medium difficulty.
The idea that algorithm sections are being pushed out of hiring is not supported by the data. An interviewing.io survey drew answers from 67 interviewers, 52 of them working at FAANG-level companies. None of those 52 reported that their company had dropped algorithm questions. But 58% changed the type of problems, about a third changed how they are presented, and 21% started giving harder ones: problems are being rewritten so that they cannot be solved by copying an answer from a chatbot. The assessment criteria and the list of topics are set out in the algorithms section breakdown.
System Design
The section is not held with everyone: it is tied to the grade and serves as a condition for moving up to a senior level. It tests the ability to clarify requirements, estimate load and justify trade-offs; knowledge of specific technologies is secondary. Hence the main mistake candidates make — starting to draw a diagram without asking a single question about what exactly is being designed and for what load.
Avito states the criteria directly: build a working architecture, weigh the pros and cons of the chosen solutions, identify and mitigate long-term risks. It also names the typical mistake — the candidate gets the task and does not know where to start the design. What exactly is asked in System Design is covered in its own breakdown.
Experience and code review
The experience section is a walk-through of a real project of the candidate’s: what problems were solved, why those particular solutions were chosen, what they would change now. There is also a reverse format, where instead of their own code the candidate is given someone else’s and asked to find the bugs in it. At Yandex the section is split into two tracks: one for those who manage people, the other for those who write code themselves. At Ozon, candidates describe a two-hour section with two interviewers in which coding, SQL and a review of someone else’s code follow one after another.
The final: team, culture fit and team matching
Technical skills are no longer tested at the final. The question here is whether the candidate will fit in with the team, and a rejection can come after every section has been passed. The reasons for rejection have nothing to do with code: one-word answers and not a single question about the product; the candidate cannot explain something complex in simple terms; working styles did not match; salary expectations diverged, and this came out only after all the sections.
The general rule of this stage: passing the technical sections does not yet guarantee an offer, because between the sections and the offer stands the agreement of a specific team. At Yandex, after the sections the candidate is shown to interested teams, and if none is ready to take them, the process stops; the company calls this team matching. In a report published in March 2026, a candidate with twenty years of experience passed the concurrency section and the algorithms section, reached two finals and was rejected. How team matching works and why candidates are rejected at the final — see the final stage breakdown.
Offer and grade
The grade is assigned from the section results; the vacancy the candidate applied for does not set it. After the meeting, the interviewer records their questions, the candidate’s answers and an assessment: the level at which to hire. The final grade is built from the assessments for all the sections, so one weak section drags down the whole result, even if the others were passed strongly. In practice it looks like this: a candidate interviewing at T-Bank got senior for system design, middle+ for the language section and junior+ for the management one.
This mechanism has two consequences. The first is downleveling at the offer stage: a developer who went through selection at Ozon for a team lead role received an offer for a senior position. The second is that there may be no vacancy at all. In a report from March 2026, a Yandex recruiter wrote to the candidate that “the colleagues who reviewed the sections set the grade at Middle+, and because of that I doubt we’ll be able to find you a relevant vacancy”; it never got as far as the team finals. The difference between grades is above all money. The offer states it before tax, and less arrives in hand: the salary calculator computes personal income tax (NDFL) on the progressive scale cumulatively over the year and shows the month in which the rate rises to 15%.
Russian company funnels side by side
| Company | Stages | Time to offer | What sets it apart |
|---|---|---|---|
| Yandex | Screening → 2–3 sections for junior and middle, 4–5 for senior → one to three finals with teams | officially ~3–4 weeks | A single assessment system by profession and level: having passed the sections once, a candidate can apply for similar vacancies in any of the company’s 90+ services without repeating the technical rounds. |
| T-Bank | Recruiter (30–60 min) → sections: language up to 90 min, algorithms 60 min, system design 60 min → meeting the team | offer a few days after the final | Detailed feedback after every section — a rare case among large companies. |
| Avito | 4–5 meetings: scoring interview 30 min → technical 2–3 hours (programming, platform, design) → final 1 hour | — | The hiring process is published as an open repository on GitHub. A retake after rejection is possible after 6 months. |
| VK | HR → technical screening → live coding → finals with teams | officially “3–4 meetings” | There can be several finals: different products consider the candidate in parallel. |
| Sber | Recruiter → technical interview, with the questions decided by the hiring manager | — | No single set of sections: according to the bank itself, its agile perimeter has more than 3,000 product teams, and the process is adapted to the project. |
| Ozon | Screening 40–50 min → 2-hour technical section with two interviewers (coding, SQL, code review) → System Design | weeks, not months | This row is compiled from candidate reports: Ozon puts its official hiring page behind anti-bot protection, and the primary source cannot be checked. |
What companies promise and what happens
Official process descriptions are consistently more optimistic than what candidates describe, and they diverge in three predictable places: timelines, feedback and the number of meetings. The shift is always in the same direction — the real process is longer and has more steps than stated.
The timeline Yandex states is 3–4 weeks for the whole cycle. A candidate who published a detailed, dated interview diary in March 2026 went from January 30 to March 4: about five weeks, and that with the algorithms section credited from a previous result. For another candidate, about six weeks passed from first contact to the final rejection. The company itself acknowledges this indirectly: in its post on the hiring update, Yandex quotes the comment “I was getting a job at Yandex, it took several months” and notes that it will surprise few people.
Yandex promises feedback a day or two after a section. In candidate reports, gaps between stages reach a week. The complaint is not about one company: in a Habr Career study of 2,100 IT specialists conducted in December 2025, 83% of respondents named the lack of feedback as the main difficulty of the interview stage.
The number of meetings diverges for a substantive reason. VK describes the process as “3–4 in-person or online meetings”, but in a report from February 2026 there were five: there can be several finals, since different products consider the candidate in parallel, and the official figure does not account for that.
A separate category of discrepancy is stages that do not exist. Some hiring materials that show up in search results describe an Advanced Code section at Yandex in detail; the company’s official pages — on the backend, frontend or mobile track — have no such section. The only reliable description of the set of sections remains the company’s own website.
AI in the interview
The market has split in opposite directions. Some companies build AI into the interview and assess the ability to work with it. Since June 2025, Canva has explicitly expected candidates for engineering roles to use Copilot, Cursor and Claude in the technical section. What it tests there is something else: whether the candidate can find and fix bugs in generated code and bring it to a state in which it can be shipped to users. Other companies take the opposite path and check the candidate through physical presence: over the past year Google and Cisco moved some interviews back to the office, citing a rise in AI cheating.
In Russia only a handful of companies publish such rules, and they are set at the track level: the requirements for a backend and a frontend developer at the same company can differ. Yandex describes its rules in the most detail. On the backend track, AI assistants are banned by name — Copilot, ChatGPT, Cursor — and the company publishes step-by-step instructions for disabling them in the IDE. The same page says the meeting may be recorded and transcribed with AI tools in order to grade the section. On the same company’s frontend track, nothing is said about AI.
How often candidates use AI, nobody knows: there are many suspicions and little evidence. In October 2025, interviewing.io surveyed interviewers — 67 responses, 52 of them from employees of FAANG-level companies. Among those 52, 81% suspect candidates of using AI. About a third have encountered it in practice. 11% have tools that detect AI, and almost all of those interviewers work at one company. The sample is small and self-selected: the respondents were users of the platform itself.
Sources
Official process descriptions are what a company states; candidate reports are how it happens in practice.
- Yandex updates its developer hiring process (in Russian) — October 2025: number of sections by grade, a single assessment across all services.
- Algorithm interviews at Yandex (in Russian) — December 2024: assessment criteria from the company itself.
- Avito’s hiring playbook (in Russian) — the set of meetings, section criteria, the ban on library functions, the cooldown before a retake.
- T-Bank: how the interview works (in Russian) — stages, section timings by track, the position on AI.
- Interviews at Yandex and VK, backend, February 2026 (in Russian) — a candidate’s dated diary: actual timelines, downleveling to Middle+, the résumé mismatch.
- A report on interviews at Ozon, T-Bank and other companies (in Russian) — July 2025: different grades for different sections, downleveling from team lead to senior.
- How is AI changing interview processes — October 2025: what happened to algorithm sections and how often companies notice AI use. The sample is small and self-selected — 67 interviewers, users of the platform itself.
- SuperJob: how Russians feel about take-home assignments (in Russian) — July 2025, 1,400 respondents: 27% of programmers do not apply to vacancies with a take-home assignment.
- SuperJob: every second Russian wants AI banned from recruitment (in Russian) — May 2026, 1,600 respondents, 379 localities.
- Habr Career: career expectations of IT specialists 2025–2026 (in Russian) — December 2025, 2,100 respondents: the main difficulties of the interview stage, including feedback.
- A recruiter’s view: 43,000 applications in a year (in Russian) — March 2026, one company’s (Infostart) annual hiring funnel: applicants per position, time spent on a résumé, the drop-out rate. Data from a single employer, not the market.
- Canva: yes, you can use AI in our interviews — June 2025, the company’s engineering blog: which tools are allowed and what is assessed instead of a ban.
- Companies bring back in-person interviews to counter AI cheating — August 2025: comments from Google and Cisco press offices that some interviews, not all, were moved back to the office.