StackShadeStackShade Library
All courses
Artificial IntelligenceFeatured path

Natural Language Processing

A complete NLP journey: text processing, linguistics, statistical models, word representations, neural sequence models, attention, Transformers, language-model pretraining, fine-tuning, retrieval-augmented generation, evaluation, safety and production systems. Every topic is taught visually, with active recall, spaced review, practical artifacts and a dedicated screen-recordable presentation.

All Levels
10 weeks
11 modules · 66 lessons
STACKSHADE / COURSE
Artificial Intelligence

Natural Language Processing

LEARNER OUTCOMES

What you should be able to do

  • Explain the full NLP stack from raw text and tokens to modern Transformer and LLM systems
  • Build classical and neural NLP pipelines and understand when each approach is appropriate
  • Reason about embeddings, attention, pretraining, fine-tuning, RAG and evaluation rather than treating them as black boxes
  • Design a production NLP system with data, retrieval, inference, evaluation, monitoring and cost trade-offs
TOPIC FORMAT

Every topic has two modes

Read — deep article with diagrams, artifacts and retrieval.

Present — focused deck with notes, pointer, timer and fullscreen.

Review — spaced review queue with due dates and recall ratings.

Focus — Pomodoro sessions, daily goals and a persistent scratchpad.

Loading progress…

Read → predict → inspect → retrieve. Your progress stays on this device.

0%
Loading study desk…
STUDY LOOP

Designed around learning, not scrolling

Each topic now has the full loop: article, visual explanation, artifact, retrieval and a recordable presentation.

Retrieval

Recall before revealing. The attempt is part of the learning action.

Spacing

Return at increasing intervals instead of rereading everything immediately.

Generation

Predict the next state or answer before seeing the worked version.

Dual coding

Pair compact prose with diagrams, traces, tables and motion.

Feynman

Teach the idea back in plain language to expose missing links.

Interleaving

Contrast nearby problem types so recognition becomes more useful.

CURRICULUM

Pick a topic

What NLP Actually Solves
Read12 min
ReadDeck
Text-to-Structure Pipeline
Interactive15 min
ReadDeck
Unicode, Normalization & Cleaning
Read12 min
ReadDeck
Tokenization: Words, Subwords & Characters
Read16 min
ReadDeck
Sentence Segmentation & Boundary Ambiguity
Read12 min
ReadDeck
Build a Reproducible Corpus Preprocessing Pipeline
Practice35 min
ReadDeck
FEYNMAN PROMPT

Explain how a messy paragraph becomes tokens and features, and why every preprocessing choice can remove information.

ACTIVE RECALL
Why is tokenization a modeling decision rather than a harmless preprocessing step?

Because the token boundary determines the units the model sees. It affects vocabulary size, sequence length, rare-word handling and what information is easy or hard for the model to represent.

What can go wrong during aggressive normalization?

Useful distinctions such as case, punctuation, accents, formatting or domain-specific symbols can disappear, hurting downstream tasks.

KEEP GOING

Another path next

View all

DSA Patterns

The 90-day pattern-based journey to Google-level DSA

Open

Computer Networks

How the Internet actually moves your data — visually

Open

System Design Fundamentals

The building blocks behind every scalable system

Open