Zero to AI Engineer

Module 53 of 54

Module 53: Glossary of AI Engineering Terms

5 min read821 words
What you'll learn
Recognize the key terms from across the pathExplain each in simple languageUse this page as a quick reference

"One page, the whole vocabulary. Once these terms feel familiar, AI articles, docs, and job posts read like plain English."

Level: Reference · Time: ~browse · Prerequisites: none

Learning Objectives

By the end of this module, you will be able to:

  • Recognize the key terms from across the path
  • Explain each in simple language
  • Use this page as a quick reference

This page gathers the vocabulary from the whole path into one place. Don't try to memorize it in a sitting — that's not how technical vocabulary sticks. Instead, skim it now to see the shape of the field, then return whenever a term feels hazy while you build. Terms cement through use: the third time you meet "embedding" in a real task, it stops being a definition and becomes an instinct. The four groups below mirror the course's four big movements, so a term's neighbors hint at when it matters.

Explain like I'm new: Learning a field is partly learning its language. Once these words feel ordinary, a dense AI blog post or job description stops looking like a wall of jargon and starts reading like a normal sentence — which is exactly when you know you belong in the conversation.

Foundations & ML

TermPlain meaning
Artificial IntelligenceMachines doing tasks that need "intelligence"
Machine learningLearning patterns from data instead of hand-written rules
Deep learningML using many-layered neural networks
Neural networkLayers of simple units ("neurons") with learnable weights
WeightsThe numbers a network adjusts as it learns
Training / inferenceLearning from data / using the trained model
Features / labelInputs (X) / the answer to predict (y)
OverfittingMemorizing noise; fails on new data
CNNNetwork great at images
TransformerArchitecture using "attention"; powers modern AI
AttentionWeighing how much each word relates to others

Generative AI & LLMs

TermPlain meaning
Generative AIAI that creates new content
LLMLarge Language Model — predicts the next token
TokenA chunk of text (~¾ word)
Context windowHow much text a model considers at once
PromptThe instruction you give a model
HallucinationConfident but false model output
EmbeddingText turned into meaning-capturing numbers
RAGRetrieve facts, then answer from them

LangChain & MCP

TermPlain meaning
LangChainFramework of building blocks for LLM apps
ChainConnected steps (prompt → model → parser)
AgentLLM that decides its own steps using tools
Tool / function callingAn action the model can request; app runs it
Vector storeDatabase of embeddings for semantic search
MCPModel Context Protocol — "USB-C for AI"
Host / client / serverThe MCP roles (app / connector / capability provider)
Resources / prompts (MCP)Read-only data / reusable templates a server offers
Prompt injectionHidden malicious instructions in read content

Try this: Pick any three terms above you couldn't have defined a month ago and say each aloud in one plain sentence, without peeking. The ones that come easily are truly yours; the ones that stall are your short revision list. That five-minute self-check is more useful than re-reading the whole table.

Edge AI & Production

TermPlain meaning
Edge AIRunning models on local devices
SLMSmall Language Model — runs on-device
QuantizationFewer bits per weight — smaller, faster
Foundry Local / runtimeSoftware that runs models locally
MLOps / SLMOpsOperating models in production / on the edge
PipelineAutomated ML steps (ingest → … → deploy)
ServingExposing a model as a live API
Feature storeConsistent features for training + serving
DriftAccuracy slipping as the world changes
CI/CD/CTAutomated test/deploy, plus continuous training
Container / KubernetesPackaged app / orchestrator for scaling services

Key idea: See the arc: foundations explain how AI learns; GenAI/LLMs are today's models; LangChain/MCP connect them to prompts, data, and tools; Edge AI runs them on devices; production ships and sustains them. That's the whole "Zero to AI Engineer" path in one map.

Notice how the terms depend on each other, which is why the arc matters more than any single row. You can't really grasp RAG without embedding and vector store; drift only makes sense once you understand training versus inference; quantization presupposes you know what weights are. A glossary looks like a flat list, but the concepts form a small web — and seeing that web is what turns scattered definitions into a working mental model of how a real AI system fits together.

Key Takeaway: These ~40 terms span the entire path — foundations and ML, generative AI and LLMs, LangChain and MCP, edge AI, and production engineering. Bookmark this page; it's your quick reference as you build.

Further Learning

Part of "Zero to AI Engineer." Terms adapted from Microsoft's open curricula (MIT License).