Research Engineer / Scientist в IT — рынок СНГ и Европы

Медиана Research — $5625/мес по данным Zorky CRM (167 активных вакансий — премиум-ниша). Research — самая высокооплачиваемая AI/ML-роль за счёт rare-skill combination (deep technical + paper-writing + novel architecture design). Senior в РФ tech (Яндекс Research / Сбер AI Foundation Models / Tinkoff Research) — $7000-12000. Senior в международных frontier-AI labs (Anthropic / OpenAI / DeepMind / Meta AI / Microsoft Research / Anthropic / Inflection / Adept / Mistral / Cohere) — $15000-30000+ /мес с equity (часто equity > base salary). Топ-выбросы: Staff / Principal Researcher в Big Tech AI — $30000-50000+, Foundation Model lead researchers — $50000-100000+ для exceptional candidates (Anthropic / OpenAI отдельные офферы 2026). PhD из top schools (Stanford / MIT / CMU / Berkeley) — премия 20-50%. Papers at top venues (NeurIPS / ICML / ICLR oral) — премия 30-100%. Citations h-index > 20 — отдельный premium-tier.

Junior Research — typical entry: 1) через MS/PhD с published papers, 2) через ML Engineer Senior + research interest + первые publications. Скачок Junior → Middle — первые first-author papers at A-tier venues (NeurIPS / ICML / ICLR). Middle → Senior — own research direction (несколько публикаций на одной теме) + invited talks + supervising Junior researchers. Senior → Staff / Principal — research agenda setting + multi-team scientific leadership + grants / funding (для academia / industry-research). Career-flow: PhD (4-6 лет) → Junior Research → Middle (2-3 года) → Senior → либо Staff / Principal Researcher (deep), либо Research Lead (management + funding), либо Founder / CTO AI-startup, либо переход в Big Tech AI с topm equity. Alternative path: ML Engineer Senior + research interest → ML Research Engineer (engineering-heavy research role) — без PhD, но требует proven novel work.

Москва Senior Research — $7000-12000/мес (Яндекс Research — крупнейший research-работодатель РФ; Сбер AI Foundation Models — GigaChat / Кандинский team; AIRI — Artificial Intelligence Research Institute; Tinkoff Research; Skoltech AI Center). СПб $6000-10000 (JetBrains Research, ИТМО AI Lab, MIPT Research). Минск/Киев Senior $5500-9500 (Sberbank AI Belarus, EPAM Research). Польша €7000-12000 gross Senior (Warsaw AI Research, IDEAS NCBR). Германия €85-130K/год Senior + bonus (Berlin AI cluster — Aleph Alpha, Tübingen — Max Planck Institute). UK Senior £70-130K (DeepMind, Cohere London). 31.3% — удалёнка. Международные frontier-AI labs (Anthropic / OpenAI / DeepMind / Meta AI / Mistral / Cohere / Inflection / Adept) — full-remote для exceptional candidates, $15000-30000+ с equity для русскоязычных Senior с английским и published research. Big Tech AI (Google Research / Microsoft Research / Apple ML / Amazon Science) — $20000-40000+ Senior+ + equity. Топ-выбросы для exceptional candidates — $50000-100000+ (Foundation Model lead researchers).

Топ-5: go, rust, python, pytorch, dynamics. Python — монопольно. PyTorch доминирует 2026 в research (80%+). JAX + Flax / Haiku — Google / DeepMind research preferred — rare premium-skill (vmap / pmap / grad — functional-style + automatic vectorization, преимущество для TPU и certain research patterns). TensorFlow — legacy, almost gone из new research 2026. CUDA kernel writing — advanced (для compute-bound research, Triton от OpenAI как Python-friendly alternative). torch.compile + JIT — optimization. Hugging Face Transformers + datasets + tokenizers — NLP / LLM research infrastructure (de-facto standard). PEFT + TRL + Accelerate (Hugging Face — fine-tuning + RLHF). Distributed training mastery: PyTorch DDP / FSDP + DeepSpeed + Megatron-LM + ColossalAI (для large-scale model training). Experiment tracking: Weights & Biases (research-favorite — best UX для experiments), MLflow, Neptune. Configuration: Hydra + OmegaConf — must для clean experiment management. PyTorch Lightning / Fabric — training framework (reduces boilerplate). Ray Train + Ray Tune — distributed training + Bayesian HPO. Optuna — Bayesian HPO standalone. DVC — data version control. LaTeX mastery — must для papers writing (Overleaf для collaboration). arXiv + OpenReview + Papers with Code — paper workflow. Mojo (new — performance-critical research, rising). Reading hardware specs: A100 / H100 / B100 — understanding compute budget. HF Hub + Kaggle + Papers with Code — datasets + benchmarks. Дополнительно для LLM Research: vLLM + TGI + SGLang (serving для evaluation), Inspect AI + lm-evaluation-harness (LLM benchmarking), TransformerLens (mechanistic interpretability).

Research Engineer — engineering-focused. Implements paper architectures, builds research infrastructure, manages compute, supports Research Scientists с experimental velocity. Often без PhD (или MS + strong engineering). Стек: PyTorch / JAX deep + distributed training mastery + CUDA / Triton + experiment infrastructure. Может авторить papers как secondary contributor, иногда first author engineering-heavy work. Зарплата 80-90% от Research Scientist на том же уровне. Research Scientist — science-focused. Formulates novel research questions, designs experiments, writes papers (first author на A-tier venues), advises Research Engineers. PhD типично required (или MS + strong publications). Стек тот же + deeper math / theory + paper-writing mastery + grant-writing (для academia). Career path academia-equivalent (Postdoc → Asst Prof tracking). ML Research Engineer (alternative title) — engineering-heavy research role без PhD, но требует proven novel work (например — implement SOTA + improve по single key metric + publish workshop paper). Career-выбор: Research Engineer если engineering-deep + interest в research без PhD-track, Research Scientist если PhD-track + interest в scientific contribution. Big Tech AI labs (DeepMind / Meta AI / Microsoft Research) обычно балансируют 1:1 или 2:3 Engineer:Scientist в team.

Reference-cycle для research 2026: 1) Reading phase — survey 20-50 papers в area (arXiv / Papers with Code / Semantic Scholar / Connected Papers tool). Notion / Obsidian / Logseq для structured notes. Reproduce 3-5 SOTA baselines (clone repos, run, validate metrics — критическая reality-check). 2) Idea generation — identify gap / limitation / unexplored combination. Discuss с supervisor / co-researchers (weekly research meetings standard). Initial hypothesis sketch. 3) Pilot experiments (small-scale, fast iteration) — implement minimal version на small dataset / small model (e.g. GPT-2 small / ResNet-18 / DistilBERT). Track в W&B. Decide hypothesis viable или kill (kill ratio в healthy lab: ~70% pilots killed). 4) Scaling phase — если pilot promising, scale up (compute + dataset + model size). Distributed training mastery required: DDP / FSDP / DeepSpeed / Megatron-LM. Compute budget tracking (часов A100 / H100). 5) Ablation phase — systematic experiments to isolate which contribution matters (typical: 5-15 ablation configs). 6) Baseline comparison — fair comparison против 3-5 strong baselines на 3-5 datasets / benchmarks. Statistical significance (3+ seeds, error bars). 7) Writing — first draft (LaTeX через Overleaf), 8 pages для NeurIPS / ICML / ICLR, 4-6 weeks типичный draft cycle. Figures professional quality (matplotlib / seaborn / plotly + careful axis labeling / legend / colorblind palette). 8) Submission — arXiv preprint first (timestamp claim), conference submission через OpenReview. Deadlines: NeurIPS май, ICML январь, ICLR октябрь. 9) Review cycle — 3-month review, rebuttal (typical: 5-10 reviewer comments per paper, response writing), final decision. Acceptance rates 20-30% top venues. 10) Camera-ready + presentation (oral если top 1% accepted, otherwise poster). Code release on GitHub + model weights on Hugging Face Hub. 11) Production transfer (industry-research) — work with ML Engineer team to integrate research into product. Often 3-6 months gap между paper publication и production deployment. 12) Follow-up work — extend, scale, apply к new domain — typical 3-5 follow-up papers from one strong initial direction. Healthy research cycle: 1-2 papers/year per researcher (single-author depth), 3-5 papers/year (collab co-author). Top researchers в frontier labs: 5-10+ papers/year при large lab co-authorship.

Да, 31.3% Research-вакансий — full-remote или гибрид, но less чем у ML Engineer / DS из-за compute-collocation (large-scale training лучше работает on-site near data-center). Frontier-AI labs (Anthropic / OpenAI / DeepMind / Meta AI / Microsoft Research) — гибрид-standard, частично remote-friendly для exceptional candidates, особенно для interpretability / theoretical work (меньше compute-bound). EU-friendly labs (Aleph Alpha Germany / Mistral France / Cohere UK) — remote-friendly. Y Combinator AI-research-startups (Adept / Inflection / Anthropic alumni-startups) — varying. Российские (Яндекс Research / Сбер AI / AIRI) — гибрид, частично remote. Аутсорсеры с research-practice (EPAM / Luxoft AI Research) — почти всегда remote. Релокант-хабы для research: США (Bay Area / NYC — major AI labs density), UK (London — DeepMind / Cohere), Канада (Toronto — Vector Institute / Mila), Германия (Tübingen — Max Planck / Berlin), Швейцария (Zurich — ETH AI Center), Сингапур (NUS / NTU AI Lab), Израиль (Tel Aviv). Английский для international research — must (вся academic literature + community + conferences на английском).

Академический research (PhD / Postdoc / Asst Prof в университете): теоретическая глубина + freedom выбирать направление + teaching load (университеты 30-50% time на courses). Publication pressure высокая (h-index — main career metric). Compute-budget ограничен grant'ами (NSF / DARPA / Horizon Europe — типичный grant $200K-2M). Salary $50-120K USD/год equivalent в US tier-1 (Stanford / MIT / Berkeley assistant professor) — гораздо ниже industry. Career-progression slow: PhD → Postdoc → Asst Prof → Tenure (10-15 лет). Plus academic freedom + own lab + intellectual prestige. Индустриальный research (DeepMind / Anthropic / OpenAI / Microsoft Research / Yandex Research): фокус на scalable engineering-feasible work + product transfer potential + larger compute budgets ($1-100M+ training runs for foundation models — недоступные в academia). Publication freedom varies: Anthropic / OpenAI / DeepMind / Meta AI / Microsoft Research поощряют publication (research-first culture), startups менее (proprietary). Salary 3-10× академического tier-1. Career-progression: Junior Researcher → Senior → Staff / Principal — 5-10 лет. Plus huge compute access + strong engineering team + faster paper-to-production transfer. Hybrid path — many top researchers split time: 60% industry + 40% academic affiliation (Microsoft Research / DeepMind allow professor-research role hybrid). Choice rule of thumb: academic если love teaching + intellectual freedom приоритет, industry если scale + compute + 5-10× salary приоритет. Researchers часто переходят academic → industry midcareer (compute + salary), редко обратно.

В топе: Яндекс Research, Сбер AI, Anthropic. Яндекс Research — крупнейший research-работодатель РФ (Foundation Models для Yandex.GPT, Search ranking research, Computer Vision для беспилотников, voice / Алиса research, embeddings research). Сбер AI Foundation Models (GigaChat — Foundation Model team; Кандинский — text-to-image; SberDevices research). AIRI (Artificial Intelligence Research Institute — independent research lab Moscow). Tinkoff Research (transaction analytics deep research, voice). Skoltech AI Center, Института ИИ МГУ / МФТИ (academic-industry hybrid). JetBrains Research. Аутсорсеры с research practice: EPAM AI Research, Luxoft AI. Frontier-AI labs (международный remote премиум): Anthropic (Claude — safety-focused frontier), OpenAI (GPT — capabilities-focused), Google DeepMind (Gemini + AlphaFold legacy), Meta AI / FAIR (LLaMA + computer vision), Microsoft Research (deep AI partnership с OpenAI + own research), Mistral AI (open-weights frontier — France), Cohere (enterprise LLM — Canada/UK), Inflection AI, Adept, Aleph Alpha (Germany), Stability AI, AI21 Labs (Israel), xAI (Musk), Tencent AI, Alibaba DAMO. Big Tech AI Research: Google Research / Apple ML / Amazon Science / IBM Research. Academic + hybrid: Stanford HAI, MIT CSAIL, Berkeley BAIR, CMU MLD, Toronto Vector Institute, Mila Montreal, Tübingen Max Planck, ETH Zurich, Oxford / Cambridge AI groups. Y Combinator AI-research-startups — sometimes premium remote для exceptional candidates.

Roadmap: 1) Math foundations strong — линейная алгебра + матанализ + теорвер + информация теория + оптимизация на уровне «понимать proofs». Книги: Strang «Introduction to Linear Algebra», Boyd «Convex Optimization» (free PDF), Cover & Thomas «Elements of Information Theory». 2) Deep Learning theory — beyond Andrew Ng. Курсы: Stanford CS231n (CV), CS224n (NLP), CS285 (Deep RL — Sergey Levine). Книги: Goodfellow «Deep Learning» (free PDF), Murphy «Probabilistic Machine Learning» (2 vols, free). 3) PyTorch deep mastery: custom Modules / Optimizers / Autograd / hooks, distributed training (DDP / FSDP), torch.compile, profiling. 4) JAX bonus (Google labs preferred — rare-skill premium): Flax / Haiku + vmap / pmap / grad mastery. 5) Reading SOTA papers — 20-50 papers в chosen area (например LLM alignment или video generation или mechanistic interpretability). Use Connected Papers + Semantic Scholar для navigation. Replicate 3-5 SOTA baselines — критическая reality-check skill. 6) LaTeX mastery — Overleaf, paper-writing patterns, figure-quality. Чтение тренируется reading 100+ papers; written — pair-writing с experienced researcher. 7) Workshop paper first goal — submit to NeurIPS workshops / ICML workshops / ICLR workshops (acceptance 40-60% vs main conf 20-30%). Establish first publication. 8) Compute access — Kaggle (free T4 / P100 + competitions), Google Colab Pro+, lambda labs, vast.ai (cloud GPUs cheap), Hugging Face Spaces (free GPU for inference). For training large models — needs lab / company access. 9) Mentorship — find supervisor / mentor (PhD program или industry research lab). Reaching out: cold-email researchers whose papers you've replicated + improved. 10) Long-term plan: PhD-track (3-6 лет full-time research) OR ML Engineer Senior → Research Engineer pivot (engineering-heavy track without PhD). Курсы РФ: МФТИ DLSchool (free YouTube), Karpov.Courses Research Track, School21 (Сбер) AI Research. International: fast.ai Practical Deep Learning Part 2 (advanced), Hugging Face NLP Course, Stanford CS25 Transformers, DeepLearning.AI Generative AI with LLMs. Books-must: «Deep Learning» Goodfellow, «Pattern Recognition and Machine Learning» Bishop, «The Elements of Statistical Learning» Hastie. Communities: arXiv-sanity-lite, Papers with Code, Twitter ML (follow top researchers), Hugging Face Discord, EleutherAI Discord, Alignment Forum. ML Engineer Senior → Junior Research Engineer — 12-24 месяца (need first workshop paper).

167 активных открытых Research-вакансий — премиум-ниша. География: EN, 🇵🇱 Польша, INT. Источники: hh.ru (Яндекс Research / Сбер AI / AIRI tickets), LinkedIn (огромный международный research-сегмент, особенно DeepMind / Meta AI / Microsoft Research / Anthropic / OpenAI), Telegram (@ml_jobs, @aijobs, @datasciencejobs, @jobsforaiml, @airesearch), специализированные борды: aijobs.net (research section), ai-jobs.net, OpenReview (jobs board), Papers with Code jobs, JobRxiv (academic), academic jobs online. Прямые careers pages: Anthropic / OpenAI / DeepMind / Meta AI / Microsoft Research / Mistral / Cohere / Aleph Alpha / Apple ML / Google Research / Amazon Science. Academic: Mathematical Job Wiki, CS Rankings, university careers pages. Twitter — many research roles posted там first. Реальный рынок шире за счёт международного remote-сегмента (frontier-AI labs hiring русскоязычных Senior researchers с published papers). Время закрытия Senior Research — 8-16 недель (longest среди ML-ниш за счёт rare-skill требований).

Senior Research владеет полным циклом frontier-research + scientific leadership. Deep math: linear algebra + matrix calculus + probability theory + information theory + optimization theory — на уровне «прочитать любой arXiv paper без math blocks». PyTorch deep mastery: custom Modules / Autograd hooks / distributed (DDP / FSDP / DeepSpeed / Megatron-LM / ColossalAI), mixed-precision training, gradient checkpointing, ZeRO optimization. JAX bonus: Flax / Haiku + vmap / pmap для exotic architectures (часто research нужно). CUDA / Triton kernel writing: advanced — для compute-bound research. Large-scale training experience: managed real 7B-70B+ parameter training runs (не на toy datasets — actual production-scale). Paper-writing mastery: first-author papers at A-tier venues (NeurIPS / ICML / ICLR / ACL / CVPR / EMNLP), citation track record. Reading speed: 5-15 papers/week sustained, gist extraction в 30-60 minutes per paper, replicate 1-2 papers/month deep. Experiment infrastructure mastery: W&B / MLflow / Hydra / OmegaConf для clean management, reproducibility (every result back to git SHA + config). LLM-specific (2026): fine-tuning через LoRA / QLoRA + RLHF (TRL library) + DPO + Constitutional AI patterns + RAG architecture deep + evaluation methodology (RAGAS / DeepEval / lm-evaluation-harness / Inspect AI) + mechanistic interpretability basics (TransformerLens). Compute budget mastery: cost analysis для training runs (A100 / H100 hours), efficient hyperparameter search (Bayesian via Optuna, not grid). Scientific writing: clear technical writing, figure quality, statistical analysis (significance + multiple seeds + error bars). Soft: supervising Junior researchers + Engineering team collaboration + technical talks at conferences + grant-writing (для academia) + paper-review (NeurIPS / ICML / ICLR reviewer typical). Английский для Senior+ ABSOLUTE MUST — research community полностью англоязычная (papers / conferences / talks). Optional bonus: open-source contributions в Hugging Face / EleutherAI / PyTorch — резко повышают market value research-recruiters.

Данные собраны автоматически из 1000+ источников — Telegram-каналов вакансий и сайтов работы СНГ и Восточной Европы (HH, Habr Career, Djinni, DOU, NoFluffJobs, JustJoin.it, Pracuj.pl и других). Парсинг работает круглосуточно, дубликаты фильтруются по описанию и URL, аномальные значения зарплат отсекаются. Подробная методология — на странице «Как работает».