AI & Machine Learning Glossary

An AI system that can take sequences of actions to complete a goal, typically by using tools (web search, code execution, database queries, API calls) and reasoning about intermediate results. Agents differ from simple LLM calls in that they can plan, execute multiple steps, and adapt their approach based on intermediate outputs. Multi-agent systems consist of multiple specialized agents that coordinate to complete complex tasks.

Anti-Money Laundering (AML) and Know Your Customer (KYC) — regulatory compliance requirements for financial institutions to verify customer identity and monitor for suspicious financial activity. AI applications in AML/KYC include transaction monitoring for suspicious patterns, customer risk scoring, identity document verification, and adverse media screening. Models must be explainable for regulatory examination.

A field of AI that enables computers to interpret and understand visual information from images and video. Modern computer vision uses convolutional neural networks (CNNs) and vision transformers (ViTs) for tasks including image classification, object detection, semantic segmentation, optical character recognition (OCR), and video analysis.

Numerical vector representations of text, images, audio, or other data that encode semantic meaning. Items with similar meaning have vectors that are close in the embedding space. Embeddings are produced by embedding models and are used as inputs to vector databases, recommendation systems, and classification models.

Artificial intelligence inference that runs on edge devices — hardware at or near the data source (sensors, cameras, industrial controllers, mobile phones) rather than in the cloud. Edge AI reduces latency, enables operation without internet connectivity, and improves privacy by keeping sensitive data on-device. It requires model optimization (quantization, pruning) to fit within edge hardware constraints.

The process of continuing to train a pre-trained foundation model on a smaller, domain-specific dataset so the model learns to produce outputs that match a specific format, style, or domain vocabulary. Fine-tuning adjusts the model weights rather than providing information at inference time.

A large AI model trained on broad data at scale that can be adapted to a wide range of downstream tasks through fine-tuning, prompting, or retrieval augmentation. Foundation models (GPT-4, Claude, Llama, Stable Diffusion) represent a paradigm shift from task-specific models to general-purpose base models that are adapted to specific use cases.

The process of transforming raw data into representations (features) that improve ML model performance. Feature engineering includes selecting relevant variables, creating derived features (ratios, aggregations, time windows), encoding categorical variables, handling missing values, and normalizing numerical features. For tabular ML problems, feature engineering often has more impact on model performance than algorithm selection.

Fast Healthcare Interoperability Resources — the HL7 standard for exchanging electronic health records (EHR) between systems. FHIR defines data formats (Resources) and APIs for clinical data including patients, observations, conditions, medications, and procedures. Compliance with FHIR R4 is required for CMS interoperability rules and is the standard integration point for EHR systems including Epic, Cerner, and Allscripts.

A capability in modern LLMs (GPT-4, Claude, Gemini) that allows the model to specify when to call a predefined function and what arguments to pass — rather than generating free-form text. The application executes the function and returns results to the model, which then generates a response incorporating the result. Function calling is the foundation for tool-using AI agents.

A phenomenon where a large language model generates text that is fluent and confident-sounding but factually incorrect or entirely fabricated. Hallucinations occur because LLMs generate the most statistically probable next token, not the most factually accurate one. Mitigation strategies include retrieval augmentation (RAG), output validation, citation enforcement, and human review for high-stakes outputs.

The leading open-source platform for ML models, datasets, and deployment infrastructure. Hugging Face hosts over 500,000 pre-trained models including BERT, T5, Llama, Stable Diffusion, and Whisper. The Transformers library provides a unified API for fine-tuning and inference. Hugging Face Inference Endpoints and Spaces provide hosting infrastructure for model deployment.

The process of running a trained ML model to produce predictions on new data. Inference is distinct from training — training adjusts model weights on a large dataset, while inference applies the fixed weights to new inputs to generate outputs. Production AI systems spend most of their compute time on inference, making inference optimization (latency, throughput, cost) a critical engineering concern.

A neural network trained on large amounts of text data that learns to predict the next token in a sequence. At sufficient scale, LLMs develop emergent capabilities including reasoning, code generation, translation, and instruction following. Modern LLMs (GPT-4, Claude, Gemini, Llama) serve as general-purpose reasoning engines for a wide range of AI applications.

An open-source framework for building LLM-powered applications. LangChain provides abstractions for chaining LLM calls with tools, memory, and retrieval systems. LangGraph extends LangChain with stateful multi-agent graph execution. LangSmith provides observability and evaluation for production LLM applications. The framework simplifies building RAG pipelines, agents, and multi-step LLM workflows.

The time delay between when an AI inference request is initiated and when the result is returned. Latency is measured as p50 (median), p95, and p99 percentile response times. Production AI applications have strict latency requirements: real-time fraud scoring under 100ms, conversational AI under 2 seconds for first token, image classification under 50ms for production line inspection. Latency optimization involves model quantization, batching, caching, and hardware selection.

Machine Learning Operations — the set of practices, tools, and workflows that enable organizations to reliably build, deploy, monitor, and maintain ML models in production. MLOps covers the full ML lifecycle: data pipelines, feature engineering, model training, evaluation, deployment, monitoring, and retraining. Analogous to DevOps for software, but with the additional complexity of managing data and model artifacts.

The degradation of a deployed ML model's performance over time as the statistical properties of real-world data diverge from the training data distribution. Data drift refers to changes in input feature distributions; concept drift refers to changes in the relationship between inputs and the target variable. Monitoring for drift and triggering retraining are core MLOps responsibilities.

The infrastructure and processes for making trained ML models available for real-time or batch inference. Model serving includes the inference server (TF Serving, Triton, FastAPI), model versioning and A/B testing, scaling and load balancing, monitoring, and optimization for latency and throughput. Production model serving is distinct from development inference and requires careful engineering for reliability.

An open-source MLOps platform for tracking ML experiments, packaging models, and managing the model lifecycle. MLflow's four components: Tracking (log parameters, metrics, and artifacts for every training run), Projects (package ML code for reproducibility), Models (a standard format for saving and loading models), and Registry (a model store with staging/production lifecycle management). Integrates with all major cloud ML platforms.

Natural Language Processing — the field of AI concerned with enabling computers to understand, interpret, and generate human language. Modern NLP is dominated by transformer-based models (BERT, GPT, T5, RoBERTa) pre-trained on large text corpora. Applications include text classification, named entity recognition, sentiment analysis, question answering, summarization, and machine translation.

Optical Character Recognition — technology that converts images of text (scanned documents, photos, screenshots) into machine-readable text. Modern AI-powered OCR (also called Intelligent Document Processing or IDP) goes beyond character recognition to extract structured data from complex document layouts including tables, forms, invoices, and contracts.

The practice of designing and optimizing the input text (prompt) sent to a language model to elicit specific, reliable, and high-quality outputs. Includes techniques such as few-shot prompting (providing examples), chain-of-thought (asking the model to reason step by step), and system prompt design (establishing role and constraints). Prompt engineering is a complement to, not a replacement for, model fine-tuning or retrieval augmentation for production systems.

A model compression technique that reduces the numerical precision of model weights from 32-bit or 16-bit floating point to 8-bit integers or lower. Quantization reduces model size (typically by 4x for INT8) and improves inference speed with minimal accuracy loss on most tasks. Essential for deploying ML models on edge devices with limited memory and compute.

A technique that combines a large language model with a retrieval system — a vector database or search index — so the model can access up-to-date, domain-specific information at inference time rather than relying solely on knowledge encoded during training. The model retrieves relevant documents, then generates an answer grounded in the retrieved content.

Reinforcement Learning from Human Feedback — a training technique used to align language models with human preferences. Human raters compare model outputs and indicate which is better; these preferences train a reward model; and the reward model guides further model training via reinforcement learning. RLHF is how models like GPT-4 and Claude are made helpful, harmless, and honest rather than just statistically likely.

Retrieval-Augmented Generation — see Retrieval-Augmented Generation for the full definition. RAG is the dominant architecture for production LLM applications that need accurate, grounded, and auditable outputs over private or frequently updated knowledge.

Software as a Medical Device — software intended to be used for one or more medical purposes without being part of a hardware medical device. The FDA, IMDRF, and EU MDR/IVDR all have specific regulatory frameworks for SaMD. AI/ML-based SaMD that continuously learns from post-market data requires a Predetermined Change Control Plan (PCCP) to receive FDA clearance.

The unit of text that language models process. Tokens are subword units — roughly 3/4 of a word on average in English. LLM API pricing, context window limits, and generation speed are all measured in tokens. Understanding token economics is critical for designing cost-efficient LLM applications at scale.

A database designed to store and query high-dimensional vector representations (embeddings). Unlike traditional databases that find exact matches, vector databases find semantically similar items using approximate nearest neighbor (ANN) algorithms. Used in RAG systems, semantic search, recommendation engines, and anomaly detection.