NLP engineering is one of the more nuanced paths in AI — sitting between classical linguistic modelling, transformer-based deep learning, and the newer LLM engineering discipline. Getting the positioning right is what determines how quickly you break in.
NLP Engineering in 2026: What Has and Hasn't Changed
The arrival of large language models has reshaped NLP significantly but not entirely. A large chunk of NLP work in production — particularly in domains like legal, healthcare, and financial services — still uses fine-tuned smaller models, carefully crafted text pipelines, and domain-specific NER (Named Entity Recognition) systems. These systems require engineers who understand the foundations: tokenisation, text preprocessing, sequence labelling, and the transformer architecture well enough to fine-tune it, not just use it via an API.
At the same time, NLP engineers are increasingly expected to understand LLM engineering patterns — RAG architectures, prompt design, and evaluation of generative outputs. The two disciplines are converging, and the most employable NLP engineers in 2026 have skills in both.
Core Technical Skills
Hugging Face Transformers is the essential starting point for modern NLP. You need to go beyond loading a pre-trained model and calling it — understand how to fine-tune BERT-family models for classification, token classification (NER), and sequence-to-sequence tasks. Know how to prepare datasets in the right format, understand training hyperparameters, and evaluate correctly.
spaCy is the production NLP workhorse. Even at companies heavily using transformer models, spaCy pipelines handle preprocessing, entity extraction, dependency parsing, and pipeline orchestration. Learn to build custom components, train spaCy models on domain-specific data, and integrate spaCy with transformer models.
Text preprocessing fundamentals: tokenisation (BPE, WordPiece, SentencePiece), handling Unicode and multilingual text, text normalisation, dealing with domain-specific vocabulary. These feel basic but are where production pipelines break.
Evaluation: Understanding the right metrics for NLP tasks — precision/recall/F1 for classification and NER, BLEU/ROUGE for generation, BERTScore for semantic similarity — and knowing their limitations. Many NLP systems fail silently in production because they're evaluated on the wrong metrics.
NLP learning path (data science route)
- Hugging Face NLP Course (free, online) — the best structured introduction to modern NLP. Work through all chapters.
- spaCy course (free at spacy.io) — production NLP fundamentals. Essential for any production NLP role.
- Stanford CS224N (NLP with Deep Learning) — covers the theoretical foundations. At minimum, watch the first 8 lectures.
- Fine-tune BERT on a domain-specific NER task — find a domain dataset (legal, medical, financial), fine-tune, evaluate properly, deploy as an API.
- Build a text classification pipeline — multi-class or multi-label, with preprocessing, training, evaluation, and serving. Document everything.
NLP vs LLM Engineering: Which Path?
If you're starting from scratch, it's worth understanding how these paths differ before committing to one. For a detailed comparison, see our NLP Engineer vs LLM Engineer guide. The short version: NLP engineers go deeper into the linguistic and model foundations; LLM engineers are primarily builders who orchestrate pre-trained models into products. Both paths are valid and the skills increasingly overlap.
The NLP path is typically preferred if you're interested in: fine-tuning models for specific domains, working in regulated industries with specific data constraints, research-adjacent roles, or building the foundational text understanding layer rather than the generative product layer.
Entry Routes by Background
From data science: The most common and typically the fastest route. You already have Python, statistical foundations, and experience working with data. The transition focuses on developing the NLP-specific stack (Transformers, spaCy), understanding transformer architecture, and adding production engineering skills. Timeline: 6–12 months.
From software engineering: Your production engineering skills transfer well. You need to add ML foundations (how models are trained, gradient descent, evaluation metrics), transformer architecture, and the NLP-specific toolchain. Timeline: 8–14 months.
From linguistics/computational linguistics: You bring rare domain knowledge and linguistic intuition. The gap is typically the engineering and ML stack. You may need to develop Python proficiency, ML fundamentals, and production deployment skills. Timeline: 6–12 months focused on the engineering layer.
Portfolio Projects That Work for NLP Roles
- Domain-specific NER system: Fine-tune a NER model on a domain dataset (legal entities, medical conditions, financial instruments). Build a processing pipeline, evaluate carefully, deploy as an API. Shows both NLP depth and engineering ability.
- Text classification at scale: Multi-class or multi-label classification on a non-trivial dataset. Handles class imbalance, multilingual text, or noisy labels. Includes a serving layer.
- Document similarity / semantic search: Build a semantic search system using sentence transformers and a vector database. Shows understanding of embeddings relevant to both NLP and LLM engineering.
See the full NLP Engineer role guide
Salary benchmarks, required skills, top UK employers, and career progression paths.
Frequently Asked Questions
Is NLP engineering different from LLM engineering?
Yes, though increasingly overlapping. NLP engineers go deeper into linguistic and model foundations — fine-tuning transformers, building text pipelines, NER, classification. LLM engineers work with pre-trained models via APIs. See our NLP vs LLM Engineer guide for detail.
What languages do NLP engineers use?
Python primarily — Hugging Face Transformers, spaCy, NLTK, PyTorch. SQL for data work. Some roles require Java/Scala for big data NLP pipelines.
Do I need a linguistics background?
No, but it helps for domain-specific roles. Most NLP engineers come from CS, maths, or data science. Strong ML foundations and NLP Python stack proficiency are what's required.
How long does it take?
From data science: 6–12 months. From software engineering: 8–14 months. From linguistics with programming skills: 6–10 months. From scratch: 2–3 years.
Which UK industries hire NLP engineers?
Legaltech, financial services, healthcare, customer service automation, media/publishing, and government. The UK's strong legaltech and fintech sectors drive particular demand.