NLP Engineer: Markt

Stelle

eine offene Stelle, die das Qualitäts-Gate passiert hat und mindestens zwei Technologien nennt

Beobachteter Kandidat

ein Kandidat, dessen Stack diese Technologie enthält; ein Aggregat — kein einziger Datensatz verlässt den Perimeter

Matchbarer Kandidat

noch nicht gezählt

Fenster

Stellen, die zum Zeitpunkt der Erstellung des Snapshots offen waren

Numbers based on a small sample — for broader benchmarks see ML Engineer and AI / LLM Engineer pages. Junior — typical entry: Computational linguistics MS / Backend Middle + NLP portfolio (Hugging Face fine-tuning experience). Junior → Middle jump — after the first production NLP feature (semantic search / sentiment classifier / NER / translation). Middle → Senior — multi-task NLP pipeline ownership + speech ASR / TTS expertise or multilingual mastery. Senior → Staff / Principal — org-wide NLP architecture + research-paper publication track. Career flow: Computational linguist / Backend Middle / DS Middle (2-3 years) + NLP interest → NLP Engineer Junior (1-2 years) → Middle (2-3 years) → Senior → either Speech Engineer specialisation, AI Engineer pivot (LLM-product focus), or Research (academic-track NLP).

Top 5: python, sql, llm, c++, git. Python monopolistic. Hugging Face Transformers mastery — single most important library 2026 (models + datasets + tokenizers + PEFT + Accelerate). spaCy (production-grade NLP pipelines — NER + POS + dependency parsing + lemmatisation — Cython-fast, industry standard for traditional NLP). NLTK (academic / legacy). Gensim (Word2Vec / LDA legacy). BERTopic (modern topic modelling — embeddings-based, rising 2024+). Modern transformers: BERT family (DeBERTa-v3 — best base for classification + NER 2026, RoBERTa, ELECTRA), T5 family (text-to-text), BART, XLM-R + mT5 + NLLB (multilingual). Sentence embeddings: sentence-transformers + BGE + E5 + Stella + jina-embeddings-v3 (top on the MTEB benchmark). LLMs for NLP: Llama 3.x / Mistral / Qwen 2.5 / Phi 3 for few-shot classification / extraction / generation. Search & ranking: Elasticsearch + OpenSearch (BM25 + dense_vector hybrid — industry standard), Vespa (Yahoo — best for production search, complex ranking), Tantivy + Meilisearch (Rust alternatives), cross-encoder rerankers (BGE Reranker / Cohere Rerank). ASR (Speech-to-Text): Whisper-large-v3 + faster-whisper (CTranslate2-optimised — production), Wav2Vec 2.0, NVIDIA NeMo, commercial APIs (AssemblyAI / Deepgram / Speechmatics / OpenAI Whisper API). Russian: Yandex SpeechKit / Sber Salute Speech / STC VoiceKit / VOSK (open-source offline). TTS: ElevenLabs (dominates 2026 — best quality), OpenAI TTS, Coqui TTS (XTTS-v2 voice cloning), Tortoise, Bark, StyleTTS 2. Russian: Yandex SpeechKit + Sber SaluteSpeech TTS. Russian NLP-specific: ruBERT + RuRoBERTa (DeepPavlov — largest Russian-language NLP project), ruGPT, FRED-T5 (Sber), ruT5, USER-bge-m3 (Russian embeddings — top on ruMTEB), Natasha (Russian NER + extraction), Razdel (Russian tokenisation), pymorphy3 (morphological analysis). Audio processing: librosa + soundfile + torchaudio. Datasets: Hugging Face Hub (>200K NLP datasets — must use), Common Crawl + OSCAR (corpus), FLORES-200 (translation benchmark), MTEB + ruMTEB (embeddings benchmarks). Evaluation: classical NLP metrics (BLEU + ROUGE + METEOR + chrF for translation, F1 + precision + recall for NER, perplexity for LM, WER for ASR, MOS for TTS), modern eval — RAGAS + DeepEval for LLM-based NLP, COMET (neural translation eval). Annotation tools: Label Studio (open-source — industry standard 2026), Prodigy (Explosion / spaCy creators), Doccano, Argilla (modern LLM-aware). Linguistic resources: WordNet, Universal Dependencies, BabelNet (multilingual).

These three roles overlap heavily in 2026 due to unification under the transformer architecture, but there are differences. ML Engineer — generalist, owns the whole production ML stack (recsys / fraud / ranking / classical ML + LLM). Stack: PyTorch + sklearn + Kubernetes + MLflow + cloud-managed ML. See ML Engineer. AI Engineer / LLM Engineer — focus on LLM integration into a product (chatbots / RAG / agents). See AI / LLM Engineer. NLP Engineer (this page) — focus on natural-language-processing tasks specifically: NER / sentiment / search / translation / summarisation / Q&A / speech (ASR + TTS). Stack overlap with AI Engineer ~60% — both use Hugging Face, embeddings, LLMs. Distinctions: 1) Classical NLP knowledge — NLP Engineer owns pre-transformer techniques (BM25 + TF-IDF + Word2Vec + LDA + dependency parsing + NER pre-BERT), AI Engineer often doesn't know this (LLM-only). 2) Speech expertise — ASR (Whisper / Wav2Vec) + TTS (ElevenLabs / Tacotron) — exclusive NLP Engineer territory (AI Engineer rarely touches speech). 3) Linguistics intuition — NLP Engineer often has a computational-linguistics background (morphology / syntax / semantics formal training), AI Engineer usually a generalist Backend / ML. 4) Multilingual / low-resource languages — NLP Engineer specialty (cross-lingual transfer, NLLB, mT5). 5) Search & ranking deep — Elasticsearch + Vespa + production ranking pipelines — NLP Engineer territory. Career pivots: NLP Engineer Senior → AI Engineer — easy lateral (1-3 months — add LangChain + agent frameworks). AI Engineer Senior → NLP Engineer — 3-6 months (classical NLP techniques + speech knowledge take time). ML Engineer Senior → NLP Engineer — 4-8 months. Reality 2026: the NLP Engineer title is giving way to AI Engineer in job postings (LLM hype), but classical NLP tasks (search / translation / extraction / speech) remain core production needs.

Yes, 17% of NLP Engineer jobs are full-remote or hybrid. NLP work is fully cloud-based (training in cloud GPUs, models in Hugging Face Hub, datasets streaming, deployment in Kubernetes). Outsourcing shops (EPAM AI / Luxoft AI / Andersen / DataArt NLP Practice) — almost always remote on US NLP projects. Russian — hybrid or remote after probation. Russian banks — hybrid/office. International voice-AI companies (ElevenLabs / AssemblyAI / Deepgram / Speechmatics / Soundhound) — full-remote standard. Big Tech NLP — hybrid-standard. English for international NLP remote — must (all NLP literature + community + conferences ACL / EMNLP / NAACL are English-speaking).

Speech Engineer — sub-specialisation within NLP focused on voice domain. Day-to-day: ASR (Speech Recognition): deploy Whisper / Wav2Vec / NeMo for transcription pipelines, fine-tune for domain-specific terminology (medical / legal / customer-support call centres — accuracy mandate), real-time streaming ASR (WebRTC + chunked processing + endpoint detection), speaker diarisation (who said what — pyannote / NVIDIA NeMo Speaker), noise robustness (denoising + voice activity detection), multilingual + code-switching support. TTS (Speech Synthesis): deploy ElevenLabs / Coqui XTTS / Tacotron / VITS for voice generation, fine-tune for brand-specific voices, voice cloning (XTTS-v2 — clone from 6 seconds of reference audio), prosody control (intonation + pacing + emotion), multilingual TTS, low-latency streaming TTS for real-time agents. Voice agents (rising 2024+): conversational AI with voice — combine ASR + LLM + TTS in a low-latency pipeline (target <500ms response). Vapi / Retell AI / Pipecat — emerging open-source frameworks. Audio processing fundamentals: librosa + soundfile + torchaudio mastery, MFCC features, spectrograms, sample-rate handling, codec knowledge (Opus / AAC / WAV). Stack-specific: NVIDIA NeMo (ASR + TTS unified framework), ESPnet (academic / research), SpeechBrain (PyTorch-based). Commercial APIs: AssemblyAI / Deepgram / Speechmatics / OpenAI Whisper API + Realtime API (Oct 2024 — voice agents) / ElevenLabs / Soundhound. Russian-specific: Yandex SpeechKit (premium for Russian — STT + TTS), Sber SaluteSpeech (banking voice + GigaChat voice), STC VoiceKit, VOSK (open-source offline). Career flow: NLP Engineer Senior + audio interest + Whisper/Wav2Vec hands-on portfolio → Speech Engineer Junior / Middle — 4-8 months.

Roadmap: 1) Linguistics fundamentals — basic understanding of morphology + syntax + semantics + pragmatics. Helps build intuition. Book: "Speech and Language Processing" Jurafsky / Martin (free online 3rd edition — bible of NLP, no need to read end-to-end, important chapters). 2) Python deep + ML basics — pandas + scikit-learn + PyTorch (basics). 3) Classical NLP — spaCy mastery + NLTK exposure. Build simple pipelines (NER + sentiment classifier + topic model with BERTopic). Course: spaCy course (free, by Explosion — spaCy creators). 4) Modern transformers fundamentals — understand BERT / DeBERTa / T5 architectures, fine-tuning workflow. Hugging Face NLP course (free, must-do — best resource 2026). 5) Hands-on Hugging Face Transformers — fine-tune DeBERTa-v3 on own classification dataset, fine-tune T5 on own summarisation task. 6) Sentence embeddings — sentence-transformers + BGE + E5. Build semantic search demo. Understand MTEB benchmark. Compare prompt-only vs fine-tuned approaches on same task. 8) Search & ranking — Elasticsearch / OpenSearch deep with hybrid (BM25 + dense). Set up production search demo. Cross-encoder reranking (BGE Reranker). 9) Speech track (optional but premium) — install Whisper / faster-whisper, build transcription pipeline + speaker diarisation (pyannote). Try ElevenLabs TTS + Coqui XTTS-v2 voice cloning. 10) Russian NLP specific (for Russian projects) — DeepPavlov framework (ruBERT + RuRoBERTa + ruGPT pre-trained models), Natasha (Russian NER), pymorphy3 (morphology), USER-bge-m3 (Russian embeddings). 11) Annotation tooling — Label Studio (industry standard 2026) — set up project + annotate small dataset + train custom model. 12) Evaluation methodology — classical metrics (BLEU / ROUGE / F1 / WER / MOS) + modern (RAGAS / DeepEval). 13) Multilingual exposure (if cross-lingual interest) — XLM-R + mT5 + NLLB-200 (Meta — 200 languages translation). 14) Pet project portfolio: a) production NER pipeline with custom domain (e.g. job descriptions extraction); b) semantic search for an open dataset; c) Russian text classification fine-tuned ruBERT; d) voice agent demo (ASR + LLM + TTS in one pipeline). Document on GitHub + blog post. Russian courses: Karpov.Courses "NLP" track, Otus "NLP", MIPT DLSchool (NLP module), SkillFactory NLP, School21 (Sber) NLP track, DeepPavlov community courses. International (EN): Hugging Face NLP Course (free, must-do), Stanford CS224N "NLP with Deep Learning" (free YouTube — best academic course), "Speech and Language Processing" Jurafsky / Martin (free PDF — bible), "Practical Natural Language Processing" Vajjala / Majumder / Gupta / Surana (O'Reilly, applied focus), fast.ai Practical Deep Learning Part 2 (NLP coverage). Must-read books: "Natural Language Processing with Transformers" Tunstall / Werra / Wolf (Hugging Face authors — must-read 2026), "Speech and Language Processing" Jurafsky / Martin. ACL / EMNLP / NAACL conferences (papers must-follow for serious NLP track). Backend Senior / DS Middle + NLP interest → NLP Engineer Junior — 4-10 months.

6 active open NLP Engineer positions with explicit NLP specifics in our sample. The real pool is many times wider — many NLP roles are classified as general ML Engineer / Backend / AI Engineer (titles like "ML Engineer for chatbot" or "Backend Engineer with NLP focus"). True NLP-focused jobs in CIS + Europe are estimated at 200–800 positions active at any moment in 2026 (counting fuzzily classified ones). Geography: the US, Poland, India. The real market is broader thanks to the international remote segment (voice-AI + NLP companies — full-remote-friendly). Time to close a Senior NLP Engineer — 6-12 weeks (longer than general AI Engineer due to rare-skill combination — linguistics + ML + audio if speech track).

A Senior NLP Engineer owns the full NLP-product engineering cycle + technical leadership. Python deep + Backend Senior level: async / typing / FastAPI / pytest mastery. Linguistics intuition: morphology + syntax + semantics + pragmatics — at the level of "I understand why the model errs in this complex case". No formal linguistics degree needed, but baseline knowledge is critical. Hugging Face Transformers mastery: models (BERT family + T5 + LLM) + datasets + tokenizers + PEFT + Accelerate. Fine-tuning mastery (LoRA / QLoRA + full fine-tuning when justified). spaCy mastery: production NLP pipelines (NER + POS + dependency parsing + custom components + matchers), spaCy-transformers integration. Modern transformers: DeBERTa-v3 (best base for classification + NER 2026), T5 family, XLM-R / mT5 / NLLB (multilingual). Understand attention + tokenisation + decoding strategies. Sentence embeddings mastery: sentence-transformers + BGE + E5 + Stella + jina-embeddings-v3, training own custom embeddings (contrastive loss + multi-negative ranking). Search & ranking mastery: Elasticsearch + OpenSearch advanced (BM25 + dense_vector hybrid + custom analyzers + multi-language support), Vespa for complex ranking pipelines, cross-encoder rerankers (training own BGE Reranker variants). LLM integration for NLP tasks: prompt engineering for NER / extraction / classification, few-shot vs fine-tuned trade-off analysis, structured output (function calling). Speech mastery (if track includes): Whisper / Wav2Vec deep (fine-tuning for domain), pyannote speaker diarisation, audio processing fundamentals (librosa + torchaudio), real-time streaming ASR architecture, voice agents architecture (ASR + LLM + TTS low-latency pipeline). Multilingual / low-resource expertise: cross-lingual transfer learning (XLM-R / mT5), data augmentation for low-resource languages, multilingual evaluation methodology. Russian NLP specifically (if Russia-focused): DeepPavlov mastery, Natasha advanced, custom Russian-specific tokenisation / morphology handling. Classical NLP knowledge: TF-IDF + BM25 internals, Word2Vec / GloVe / FastText, dependency parsing algorithms, CRF for sequence labelling — for understanding when classical beats LLM (cost / latency / explainability). Annotation tooling mastery: Label Studio advanced + Argilla (modern LLM-aware) + Prodigy (spaCy ecosystem). Evaluation mastery: classical metrics (BLEU + ROUGE + METEOR + chrF + F1 + perplexity + WER + MOS), modern (RAGAS + DeepEval + COMET for translation), human eval methodology design. System design for NLP products: design NLP pipeline on a whiteboard under scale (100M+ texts/day), latency budgets (target P95 for real-time NLP), cost optimisation (cache + batch + smart routing). Soft: ADRs writing for NLP architecture decisions, technical writing (NLP feature design docs + evaluation reports), cross-team collaboration (Product / Backend / DS / Linguists teams), mentoring Middle NLP engineers, paper-reading discipline (ACL / EMNLP / NAACL / Interspeech if speech). English for Senior+ MUST — NLP community / docs / papers / conferences ACL / EMNLP / NAACL are English-speaking. Optional bonus: open-source contributions to Hugging Face / spaCy / DeepPavlov / sentence-transformers — sharply increase market value.