Natural Language Processing Engineering: Zero to Hero

The NLP pipeline: from raw text to a served prediction

Text cleaning, normalization, and unicode handling

Regex-based extraction and rule-based text processing

Word, character, and subword tokenization tradeoffs

BPE, WordPiece, and SentencePiece in practice

Bag-of-words and n-gram features

TF-IDF weighting and building a strong classical baseline

Word2Vec and GloVe: dense semantic representations

FastText and subword-aware embeddings

N-gram language models and smoothing

Perplexity and evaluating language models

Recurrent neural networks and the vanishing gradient problem

LSTMs and GRUs in practice: sequence tagging and classification

Multi-head attention and positional information

Encoder and decoder blocks: assembling the full architecture

Training dynamics, compute scaling, and architectural variants

BERT and masked language modeling

Transfer learning: why pretrain-then-adapt beats training from scratch

Adding a task head: classification and sequence labeling

Diagnosing fine-tuning failures

BIO tagging and entity-level evaluation

Production NER: PII detection, schema evolution, and monitoring

Causal language modeling and the GPT architecture

Decoding strategies: greedy, beam search, and sampling

Multi-label, multi-class, and hierarchical classification

Aspect-based sentiment analysis and evaluation beyond accuracy

Extractive question answering: span prediction

Generative QA and multi-hop reasoning

Extractive and abstractive summarization

ROUGE, its limitations, and handling long documents

Neural machine translation and BLEU

Building genuinely multilingual and low-resource NLP systems

Sentence embeddings and contrastive training

Hybrid sparse-plus-dense retrieval

Approximate nearest neighbor search: HNSW and IVF

Operating a vector database in production

Advanced RAG patterns: re-ranking, query rewriting, and agentic retrieval

Zero-shot, few-shot, and chain-of-thought prompting

Prompt versioning, testing, and production prompt management

LoRA: low-rank adaptation

Adapters, prefix tuning, and choosing a PEFT method

Instruction tuning: from next-token predictor to helpful assistant

RLHF, preference-based alignment, and alignment failure modes

Building a multi-signal evaluation harness

Systematic hallucination detection and factuality evaluation

Dialogue state and multi-turn conversation management

Evaluating conversation quality across a full session

Measuring and mitigating bias in language models

PII detection, redaction, and content moderation guardrails

Latency, batching, and streaming for generative serving

Caching, model selection routing, and cost-aware NLP serving architecture

Enterprise NLP case studies: search, support, and document intelligence

NLP engineering interview preparation

Capstone I: Enterprise Document Intelligence Platform

Capstone II: Production RAG Assistant Defense & Evaluation Harness