Career roadmap
NLP Engineer
Build systems that read, classify and generate language, and know when a fine-tuned small model beats a large one.
Before you start NLP
- Python and PyTorch basics
- Linear algebra and probability fundamentals
- Familiarity with the transformer architecture
Language and representation
Understand how text becomes numbers, because most NLP bugs live there.
Unglamorous, and the source of most silent quality problems.
- Normalisation, Unicode and encoding issues
- Language detection and segmentation
- Handling noisy real-world text
- Regular expressions and their limits
Subword tokenisation explains cost, context limits and multilingual weakness.
- BPE, WordPiece and SentencePiece
- Vocabulary size trade-offs
- Token counts and non-English penalties
- Special tokens and truncation strategy
The representation underlying search, clustering and retrieval.
Ch — Embeddings & Vector Search- Static versus contextual embeddings
- Sentence embeddings and pooling
- Similarity metrics and normalisation
- Choosing an embedding model
Attention mechanics get asked directly in almost every NLP interview.
- Self-attention and multi-head attention
- Positional encoding schemes
- Encoder, decoder and encoder-decoder models
- Context length and attention cost
Often the right answer for cost and latency, and a differentiator in interviews.
- TF-IDF and linear classifiers as baselines
- Named entity recognition approaches
- Rule-based extraction where appropriate
- When simple methods win
BuildBuild a text classifier three ways — bag of words, embeddings, fine-tuned transformer — and compare honestly.
Core NLP tasks
The task types that make up almost all commercial NLP work.
The most common commercial NLP task by a wide margin.
- Single and multi-label classification
- Class imbalance in text
- Threshold selection and confidence
- Weak supervision and labelling functions
Turning documents into structured records is where the money usually is.
- Entity recognition and linking
- Relation extraction
- Structured output from LLMs
- Evaluating span-level accuracy
Retrieval quality determines RAG quality more than any model choice.
Ch — RAG- Lexical search and BM25
- Dense retrieval and hybrid search
- Reranking with cross-encoders
- Chunking strategies and their impact
Summarisation and rewriting, with the evaluation problems they bring.
- Summarisation approaches and failure modes
- Controlled generation and constraints
- Hallucination and grounding
- Decoding parameters
A large practical gap, and a strong differentiator in many markets.
- Cross-lingual models and transfer
- Tokenisation penalties by script
- Low-resource language strategies
- Evaluation across languages
BuildShip a document extraction pipeline with a measured accuracy figure and an error analysis.
Training and adaptation
Knowing when to prompt, when to fine-tune, and when to train something small.
Parameter-efficient tuning made this practical on modest hardware.
Ch — Prompt Engineering- Full fine-tuning versus LoRA and adapters
- Dataset construction and formatting
- Hyperparameters that actually matter
- Catastrophic forgetting
Data quality dominates model choice for task-specific performance.
- Annotation guidelines and agreement
- Active learning for labelling budgets
- Synthetic data generation and its risks
- Deduplication and contamination checks
How you get production latency and cost down without losing much quality.
- Knowledge distillation from a large model
- Quantisation and its quality cost
- Pruning and small model selection
- Latency and throughput trade-offs
Often the right first answer. Knowing when it stops being right is the skill.
- Few-shot and structured prompting
- Output schemas and validation
- Cost and latency of prompting versus tuning
- Prompt versioning and regression
The hardest part of NLP, and the part interviews probe most for rigour.
Ch — Evaluation & Hallucination- Task-appropriate metrics
- Human evaluation design
- LLM-as-judge and its biases
- Held-out sets and contamination
BuildFine-tune a small model to beat a large API model on one narrow task, on both cost and quality.
Production NLP
Serving text models under latency, cost and safety constraints.
Throughput and latency engineering for transformer inference.
- Batching and dynamic batching
- GPU versus CPU inference economics
- ONNX and inference runtimes
- Autoscaling for spiky traffic
Text pipelines have many stages; knowing which dominates is essential.
Ch — Cost & Latency- Profiling a multi-stage pipeline
- Caching embeddings and results
- Model cascades and routing
- Streaming responses
Text distributions shift, and accuracy degrades silently.
- Input distribution monitoring
- Output quality sampling
- Feedback capture from users
- Retraining triggers
User-facing language systems attract adversarial input immediately.
Ch — AI Security- Prompt injection in text pipelines
- PII detection and redaction
- Toxicity and content filtering
- Abuse rate limiting
Text is full of personal data, which brings regulation into scope quickly.
- PII handling in training data
- Data retention for prompts and outputs
- On-premises versus API trade-offs
- Consent and purpose limitation
BuildDeploy a model behind an API with monitoring, caching and a documented latency budget.
Interview preparation
NLP interviews combine architecture depth, task design and practical judgement.
Attention, positional encoding and tokenisation, asked precisely.
- Explain self-attention and its complexity
- Why positional encoding is needed
- Encoder versus decoder model choice
- Context length trade-offs
Design a search or extraction system with quality and cost targets.
- Retrieval architecture and reranking
- Evaluation plan before building
- Cost and latency budgeting
- Handling failure and fallback
Practical text processing and model usage rather than algorithm puzzles.
- Implementing a retrieval pipeline
- Writing an evaluation harness
- Data cleaning under time pressure
- PyTorch fundamentals
When would you not use an LLM is now a standard question.
- Fine-tune versus prompt versus classical
- Diagnosing a quality regression
- Building an evaluation set from scratch
- Explaining trade-offs to product
Measured results with error analysis beat model zoo demos.
- A task with a real accuracy number
- Error analysis and failure taxonomy
- A distillation or cost reduction case study
- Published evaluation methodology
BuildA public project with an evaluation report and an honest error analysis.
NLP tools on your CV
- PyTorch
- Hugging Face Transformers
- spaCy
- sentence-transformers
- FAISS / pgvector
- PEFT / LoRA
- Weights & Biases
What NLP employers ask to see
- A text task with a measured accuracy figure and error analysis
- A fine-tuned model that beats an API model on cost and quality
- A retrieval system with a documented evaluation set
- A published evaluation methodology for a language task
Absorbed into the LLM boom but still distinct: search, extraction, classification and multilingual work need people who understand text beyond calling an API.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.