Why These Terms Matter

AI engineering has moved from research labs into mainstream engineering practice faster than almost any technology before it. Structural, electrical, mechanical, and civil engineers now routinely encounter AI terminology in vendor documentation, procurement specs, IT policies, and their own firm's tool evaluations — often without a background in computer science. Unlike the NEC or ASHRAE standards, AI terminology is not codified in a single reference document, and vendors frequently use loose or overlapping language ("AI," "machine learning," "generative AI," and "agent" are often used interchangeably when they describe distinct things).

Understanding these terms precisely matters for practical reasons: it determines what questions to ask a software vendor, what a firm's AI use policy should actually restrict, whether a tool's output needs professional review, and how to evaluate claims about accuracy and data privacy. This glossary covers 55 of the most important AI engineering terms an engineer is likely to encounter, organized alphabetically with plain-language explanations and, where relevant, notes on why the distinction matters in professional practice.

A

Agent (AI Agent) — Autonomous task execution
A system built on top of a language model that can take multiple steps toward a goal without a human directing each step — calling tools, reading results, and deciding what to do next in a loop. A chatbot answers one question at a time; an agent might be told "check these three drawings for dimension conflicts" and then autonomously open each file, compare values, and compile a report. The key engineering concern is that agents can take actions (not just generate text), so error containment and human review checkpoints matter more than with a simple chat interface.
AI Governance — Organizational risk controls
The policies, review processes, and accountability structures a firm puts in place to control how AI tools are used, particularly for deliverables that carry professional liability. Good AI governance defines which tasks AI may assist with, who reviews AI-assisted output before it's stamped or submitted, and how client data may or may not be shared with AI tools.
Algorithm — Step-by-step procedure
A defined sequence of steps for solving a problem or performing a computation. In machine learning, "algorithm" often refers to the training method (e.g., gradient descent) rather than the resulting model — the algorithm is the recipe, the model is the finished dish.
API (Application Programming Interface) — Programmatic access point
A defined way for one piece of software to call another. An "AI API" (like the Claude API or OpenAI API) lets a developer send text to a model and get a response back programmatically, without using a chat website — this is how AI gets embedded into other software, such as a CAD plugin or a custom internal tool.
Artificial General Intelligence (AGI) — Hypothetical human-level AI
A hypothetical future AI system with human-level (or greater) reasoning and learning ability across essentially any task, as opposed to today's systems, which are broad but still fall short of reliable general reasoning in many domains. AGI does not exist yet; the term is useful mainly for distinguishing marketing hype from the actual capabilities of current tools.
Attention Mechanism — Core transformer component
The mathematical mechanism inside a transformer model that lets it weigh how relevant every other word in the input is to the word currently being processed. Attention is why modern language models handle long-range context (referring back to something mentioned many paragraphs earlier) far better than older architectures.

B

Bias (Model Bias) — Systematic skew in output
A systematic tendency of a model to produce skewed, unfair, or inaccurate results for certain inputs, usually traceable to patterns (or gaps) in its training data. In engineering contexts, this can show up as a model being more familiar with US codes than international ones, or defaulting to outdated code editions, simply because that's what appeared more often in its training data.
Bias-Variance Tradeoff — Model fitting balance
A core concept in traditional machine learning describing the tension between a model being too simple to capture the real pattern (high bias, underfitting) and being so complex it memorizes noise in the training data instead of the underlying trend (high variance, overfitting).

C

Chain of Thought (CoT) — Step-by-step reasoning
A prompting technique, and increasingly a built-in model behavior, where a model works through a problem in explicit intermediate steps before giving a final answer, rather than jumping straight to a conclusion. For calculation-heavy engineering tasks, chain-of-thought output is valuable because it lets a reviewer check the model's intermediate steps, not just trust the final number.
Classification — Category prediction task
A machine learning task where the model assigns an input to one of a fixed set of categories — for example, classifying a photo of a structural connection as "acceptable" or "deficient." Contrast with regression, which predicts a continuous number.
Context Window — Model's working memory
The maximum amount of text (measured in tokens) a model can consider at once, including both the prompt and its own response. A model with a 200,000-token context window can hold roughly a few hundred pages of text in a single conversation; content beyond the window is simply not visible to the model, which is why very long documents sometimes need to be split into sections.
Convolutional Neural Network (CNN) — Image-processing architecture
A neural network architecture specialized for processing grid-like data such as images, using small filters that scan across the input to detect patterns like edges, textures, and shapes. CNNs are the workhorse behind computer-vision applications such as automated defect detection in structural inspection photos.

D

Deep Learning — Multi-layer neural networks
A subset of machine learning that uses neural networks with many stacked layers to learn increasingly abstract representations of data. "Deep" refers to the number of layers, not the sophistication of the reasoning — depth is what allows the network to learn hierarchical features (edges, then shapes, then objects, for example).
Diffusion Model — Image-generation architecture
A generative model architecture that creates images (or other content) by starting from random noise and iteratively refining it into a coherent output, guided by a text prompt. Diffusion models are the technology behind most AI image generators; they are a different architecture from the transformer models used for text.

E

Embedding — Numeric meaning representation
A representation of text (or an image) as a list of numbers (a vector) positioned in a high-dimensional space such that items with similar meaning end up near each other. Embeddings are what make semantic search possible — a search for "beam deflection limits" can retrieve a document about "allowable sag in structural members" even though the wording doesn't match, because the embeddings are close together.
Epoch — Training data pass
One complete pass through the entire training dataset during model training. Training typically runs for multiple epochs, with the model's internal parameters adjusted a little after each pass.

F

Fine-Tuning — Additional specialized training
The process of taking an already-trained (pretrained) model and continuing its training on a smaller, specialized dataset to adapt it to a particular domain or task — for example, fine-tuning a general language model on a firm's past specifications to better match its writing style and terminology. Fine-tuning changes the model's internal weights, distinguishing it from prompting or RAG, which leave the model unchanged and instead supply context at query time.
Foundation Model — Large general-purpose base model
A large model trained on broad, general data (text, code, images) that serves as the base for many downstream applications, either used directly or adapted through fine-tuning. GPT, Claude, and Gemini are all foundation models.

G

Generative AI — Content-creating AI
AI systems that create new content — text, images, code, audio — rather than only classifying or predicting from existing categories. This is the category that includes chatbots and image generators, and it's distinct from earlier "predictive" AI used for things like equipment failure forecasting.
GPU (Graphics Processing Unit) — Parallel-computation hardware
A processor originally designed for rendering graphics that turns out to be extremely well-suited to the massive parallel matrix multiplication that neural networks require, making GPUs (and specialized variants like TPUs) the standard hardware for both training and running AI models.
Guardrails — Output-constraining safeguards
Rules, filters, or checks layered around a model to constrain its behavior — blocking certain topics, validating outputs against expected formats, or flagging low-confidence answers for human review. Guardrails are especially important in engineering applications where an ungrounded or out-of-scope answer could be mistaken for a validated calculation.

H

Hallucination — Fabricated but plausible output
When a model generates output that is factually wrong, fabricated, or unsupported by any real source, but stated with the same confident, fluent tone as accurate output. Hallucination is a fundamental characteristic of how generative language models work (they generate statistically plausible text, not verified facts), which is exactly why AI-assisted calculations and code citations must be independently verified before use in a stamped deliverable.
Hyperparameter — Pre-training configuration setting
A configuration value set before training begins that controls how training happens — such as learning rate or batch size — as opposed to the model's weights, which are learned during training itself.

I

Inference — Running a trained model
The process of using an already-trained model to generate a prediction or response for new input, as opposed to training, which is the (far more computationally expensive) process of creating the model in the first place. Every time an engineer submits a prompt to a chatbot, that's an inference request.
Instruction Tuning — Command-following training
A specific type of fine-tuning where a model is trained on examples of instructions paired with correct responses, teaching it to follow direct commands ("summarize this," "convert this to a table") rather than simply continuing text in the style of its training data.

K

Knowledge Base (for AI) — Curated reference document set
A curated collection of documents (specifications, code sections, past project reports) that an AI system can search and reference, typically as part of a RAG pipeline, so its answers are grounded in specific, verifiable source material rather than only its general training.

L

Large Language Model (LLM) — Text-generating neural network
A neural network, typically built on the transformer architecture and trained on enormous amounts of text, that predicts and generates natural language. LLMs are the technology behind chatbots like Claude, ChatGPT, and Gemini; the "large" refers to both the size of the training data and the number of internal parameters (often in the tens or hundreds of billions).
Latency — Response time delay
The time between sending a request to a model and receiving its response. Latency matters for real-time applications (voice assistants, live coding tools) and is a key factor engineers weigh when choosing a smaller, faster model over a larger, more capable but slower one.
LoRA (Low-Rank Adaptation) — Efficient fine-tuning method
A fine-tuning technique that trains a small number of additional parameters rather than adjusting a model's entire weight set, making customization dramatically cheaper and faster while achieving results close to full fine-tuning for many tasks.

M

Machine Learning (ML) — Learning from data
A broad field of computer science in which systems improve at a task by learning patterns from data, rather than following explicitly hand-coded rules. Machine learning is the parent category that includes deep learning, and deep learning is the parent category that includes today's large language models.
Multimodal Model — Multiple input-type model
A model capable of processing more than one type of input — text and images together, for example — allowing it to, say, read a scanned drawing and answer questions about the dimensions shown on it in the same conversation.

N

Natural Language Processing (NLP) — Language-understanding field
The broader field of AI concerned with enabling computers to understand, interpret, and generate human language; large language models are the current dominant approach to NLP, though the field predates them by decades.
Neural Network — Layered computation graph
A computational model loosely inspired by biological neurons, composed of layers of interconnected nodes ("neurons") that transform input data through weighted connections, adjusted during training to make the network's output match the desired result.

O

Overfitting — Memorizing rather than generalizing
When a model learns the training data too specifically — including its noise and quirks — so well that it performs poorly on new, unseen data. An overfit model has effectively memorized answers rather than learned the underlying pattern.

P

Parameters — Learned model values
The internal numeric values (weights and biases) a model learns during training and uses to compute its outputs. A model described as having "70 billion parameters" has that many individually adjustable values, which is one (imperfect) proxy for model capacity.
Prompt — Model input text
The text input given to a language model to elicit a response — a question, instruction, or piece of context. A "system prompt" is a special instruction set separately from the user's message, typically defining the model's role or constraints for the entire conversation.
Prompt Engineering — Crafting effective inputs
The practice of designing and refining prompts to reliably get accurate, well-formatted, useful output from a language model — including techniques like providing examples, specifying an output format, and breaking a complex task into smaller steps.
Pretraining — Initial broad-scale training
The initial, extremely resource-intensive phase of training a foundation model on a massive, general dataset, before any task-specific fine-tuning happens. Pretraining is what most people mean when they picture "training an AI model" at the scale done by major AI labs.

Q

Quantization — Reduced-precision compression
A technique that reduces the numeric precision of a model's parameters (e.g., from 32-bit to 8-bit numbers) to shrink its memory footprint and speed up inference, usually with a small, controlled tradeoff in accuracy — this is a key technique for running AI models on smaller devices or at lower cost.

R

RAG (Retrieval-Augmented Generation) — Grounded document lookup
An architecture where a model, before answering, first retrieves relevant passages from an external knowledge base (via embeddings and vector search) and includes them in its prompt, so its answer is grounded in specific source documents rather than only its general training. RAG is one of the most practical tools for reducing hallucination in engineering applications, because the model's answer can be traced back to a specific retrieved code section or specification.
Regression (Statistical/ML) — Continuous value prediction
A machine learning task where the model predicts a continuous numeric value — such as predicted energy consumption or remaining equipment life — as opposed to classification, which predicts a discrete category.
Reinforcement Learning from Human Feedback (RLHF) — Human-guided alignment training
A training technique where human reviewers rate or rank a model's candidate outputs, and that feedback is used to further train the model to prefer responses humans judge as more helpful, accurate, or appropriate. RLHF is a major part of why modern chatbots are more conversational and better at following instructions than earlier raw language models.

S

Semantic Search — Meaning-based document retrieval
A search method that retrieves results based on conceptual similarity of meaning (via embeddings) rather than exact keyword matching, allowing a query about "voltage drop limits" to surface a document that discusses "acceptable conductor loss" even without shared keywords.
Supervised Learning — Labeled-data training
A machine learning approach where a model is trained on data that includes correct, labeled answers (e.g., photos of cracks labeled "structural" or "cosmetic"), and the model learns to reproduce that labeling on new, unlabeled data.
System Prompt — Persistent instruction context
A set of instructions provided to a model separately from the ongoing conversation, typically defining its role, tone, constraints, or available tools, and persisting across the whole session rather than being just one more user message.

T

Temperature — Output randomness setting
A parameter that controls how random or deterministic a model's output is: low temperature produces more predictable, conservative responses, while high temperature produces more varied, creative (and less reliable) ones. For calculation-style or code-reference tasks, a low temperature setting is generally preferred.
Token — Basic text processing unit
The basic unit of text a language model processes — often a word, part of a word, or punctuation mark. A rough rule of thumb is that a token is about four characters of English text on average, which is why context-window limits and API pricing are both measured in tokens rather than words.
Tool Use / Function Calling — Model-invoked external actions
The ability of a model to recognize when it needs to invoke an external tool (a calculator, a database lookup, an API, a search engine) to complete a task, call it with the right parameters, and incorporate the result into its response. Tool use is the mechanism that turns a plain chatbot into an agent capable of taking real actions.
Training Data — Source learning material
The dataset used to teach a model during training — for a foundation model, typically a vast mixture of text scraped from books, websites, code repositories, and other sources, whose composition significantly shapes what the model knows and how it's biased.
Transfer Learning — Reusing prior training
The practice of taking a model already trained on one task or dataset and adapting it to a new, related task, leveraging the general patterns it already learned rather than training a new model from scratch — fine-tuning is one specific form of transfer learning.
Transformer — Dominant modern AI architecture
The neural network architecture, introduced in 2017, built around the attention mechanism, that underlies nearly all modern large language models and many image and multimodal models. The transformer's ability to process all parts of an input in parallel (rather than sequentially, as older architectures did) is a major reason for the rapid capability gains in AI since its introduction.

U

Unsupervised Learning — Pattern-finding without labels
A machine learning approach where a model finds patterns or structure in data that has no labeled correct answers — clustering similar items together, for instance — as opposed to supervised learning, which relies on labeled examples.

V

Vector Database — Embedding storage and search system
A specialized database designed to store embeddings and quickly search for the ones most similar to a given query vector, forming the retrieval backbone of most RAG systems used to ground AI answers in a firm's own documents.

W

Weights — Learned connection strengths
The numeric values assigned to the connections between neurons in a neural network, adjusted during training to minimize the difference between the model's predictions and the correct answers. "Weights" and "parameters" are often used interchangeably, though parameters technically also include bias terms.

Z

Zero-Shot Learning — No-example task performance
A model's ability to correctly perform a task it was never explicitly trained or given examples for, relying only on general knowledge and instructions in the prompt — as opposed to "few-shot," where a small number of examples are provided in the prompt to guide the model's response format or approach.