Skip to main content

NLP Engineer Roadmap 2026

Build language systems from word vectors to transformers, pretraining, fine-tuning, evaluation and agents, following the CS224N arc.

4 stages · 8 units · Advanced · about 150 hours · Last reviewed 19 November 2026

What is a nlp engineer?

An NLP engineer builds systems that read, classify, translate, summarise and generate human language. The work runs from tokenisation and word vectors to transformers, pretraining, fine-tuning, evaluation and tool-using agents. This roadmap follows the arc of Stanford's CS224N course, using Python and PyTorch for the practical parts.

Step-by-step nlp engineer roadmap

  1. 1

    Language Basics

    Text as data.

    • Text Processing

      Text Processing is the stage that turns raw language into numbers a model can use. It covers cleaning text, splitting it into tokens with schemes such as BPE or WordPiece, and representing words as dense vectors like Word2Vec or GloVe, where similar meanings sit close together in vector space.

    • Neural Foundations

      Neural Foundations are the deep learning building blocks used before transformers took over language modelling. They include feed-forward networks, backpropagation, recurrent networks such as LSTMs and GRUs, and sequence-to-sequence models with an encoder and decoder. These ideas explain why attention was invented.

  2. 2

    Transformers

    The core architecture.

    • Architecture

      Architecture in this roadmap means the transformer: a network that uses self-attention to let every token weigh every other token in a sequence. Queries, keys and values compute the weights, multiple heads capture different relations, and positional information plus residual connections and layer normalisation make deep stacks trainable.

    • Pretraining

      Pretraining is the phase where a language model learns general language ability from large unlabelled text by predicting tokens. Masked language modelling gave BERT-style encoders, and next-token prediction gave GPT-style decoders. Fine-tuning then adapts the pretrained weights to a specific task with a much smaller labelled dataset.

  3. 3

    Adaptation and Alignment

    Make models useful.

    • Post-training

      Post-training is the work done after pretraining to make a model follow instructions and behave as intended. It includes supervised instruction tuning, preference-based alignment such as RLHF or DPO, and parameter-efficient adaptation like LoRA, which trains small low-rank matrices instead of updating every weight.

    • Tasks

      Tasks are the standard problems NLP systems are built to solve, each with its own output format and metrics. They include text classification, named entity recognition, question answering, summarisation, machine translation and semantic similarity. Choosing the task framing decides the model head, the data labels and the evaluation method.

  4. 4

    Systems and Responsibility

    Build and judge.

    • Applications

      Applications are the systems built around language models rather than the models themselves. Retrieval-augmented generation fetches relevant documents with embeddings and a vector index, then feeds them to the model. Agents go further by letting a model call tools, read results and plan multi-step actions.

    • Quality

      Quality in NLP is the discipline of measuring whether a system works and whether it is fair. It covers automatic metrics such as accuracy, F1, BLEU, ROUGE and perplexity, human and model-based judging for open-ended output, and ethics checks for bias, privacy, safety and misuse before release.

Frequently asked questions

Ready to practise for interviews?

Practise nlp engineer questions with an AI mock interviewer and get feedback on your answers.

More roadmaps