AI, Computer Vision and IoT Glossary

AI and language models

AI agent

Software that uses a language model in a loop of planning, calling tools (APIs, databases, files) and evaluating results to reach a goal. Instead of producing one answer it executes several steps itself, which makes authorisation and human approval critical.

Artificial intelligence (AI)

The umbrella term for methods that let machines perform tasks that normally require human intelligence, such as seeing, understanding language, deciding and learning. In organisations it is used mostly as machine learning, deep learning and large language models.

Embedding

A representation that expresses the meaning of a text, image or other data as a vector of numbers. Semantically similar content produces nearby vectors, which makes embeddings the basis of semantic search and RAG.

Large language model (LLM)

An AI model trained on very large text corpora that generates, summarises, translates and answers questions by predicting the next token. On its own it has no current or company-specific knowledge, so it is complemented with RAG and tool calling.

MCP (Model Context Protocol)

An open protocol that lets AI applications connect to tools, data sources and prompts through a standard interface. Once a system is exposed as an MCP server, any AI client that supports the protocol can use it.

Model drift

A drop in model accuracy because field data gradually diverges from the training data. Seasons, lighting, clothing and layout changes are typical causes for vision models; it is managed with continuous measurement and regular retraining.

Quantization (INT8)

Running a model's weights and computations in 16-bit floats or 8-bit integers instead of 32-bit floats. It cuts memory and latency significantly; the accuracy loss is measured and bounded with a representative calibration dataset.

RAG (retrieval-augmented generation)

A method where relevant documents are retrieved by a search layer and passed to the language model as context before it answers. The model grounds its answer in those documents and can cite them — the most common way to use company knowledge without retraining the model.

Vector database

A database or database extension that stores embedding vectors and quickly finds the ones closest to a query vector (approximate nearest-neighbour search). In RAG systems it is where the document chunks live.

Computer vision

The field of AI that extracts meaningful information — objects, people, motion and states — from images and video. It turns camera footage into measurable outputs such as counts, alarms and metrics.

Edge AI

Running an AI model where the data is produced — the store, the factory, the device — instead of sending data to the cloud. It reduces bandwidth and latency and strengthens privacy by keeping footage on site.

Multi-object tracking (MOT)

Giving objects detected in consecutive frames a persistent identity so their movement can be followed. Without tracking the same person is counted in every frame; metrics such as counts, dwell time and line crossing depend on it.

Object detection

The computer-vision task of finding what objects are in an image (class) and where they are (bounding box). It is the first step of applications such as people counting, PPE checks and vehicle detection.

OEE (Overall Equipment Effectiveness)

An efficiency measure of production equipment calculated as availability × performance × quality. Camera analytics can be added to PLC data to understand the causes of downtime.

ONNX

An open model format that stores models trained in different frameworks (PyTorch, TensorFlow, etc.) in a common representation. With ONNX Runtime the same model can run on CPUs, GPUs and accelerators.

ONVIF

An industry standard and forum that ensures interoperability between IP-based security products such as cameras and NVRs. ONVIF-compliant cameras can be discovered and streamed the same way regardless of brand.

PPE detection

Automatically checking from camera footage whether personal protective equipment (helmet, vest, goggles, etc.) is worn. Used in occupational safety for 24/7 inspection and real-time alerts.

RTSP (Real Time Streaming Protocol)

The network protocol used to pull live video streams from IP cameras. Video-analytics systems connect to the camera or NVR over RTSP and receive the H.264/H.265 stream.

TensorRT

NVIDIA's software development kit that optimises deep-learning models for a specific GPU and compiles them into an accelerated inference engine. With FP16 and INT8 precision it markedly reduces latency.

YOLO (You Only Look Once)

A family of single-stage deep-learning models that detect objects in real time by processing the image in a single pass. Its balance of speed and accuracy makes it one of the most widely used detection architectures in the field.

IoT and networking

IoT (Internet of Things)

Sensors, machines and devices exchanging data over a network to talk to each other and to software. In industry it turns measurements into real-time events and decisions.

MQTT

A lightweight messaging protocol based on the publish/subscribe model. Devices send messages under topics to a broker; it is one of the most widely used IoT protocols.

QoS (MQTT quality of service)

The level that sets a message's delivery guarantee in MQTT: 0 at most once, 1 at least once, 2 exactly once. Higher levels mean safer but costlier communication.

SD-WAN

Managing the wide-area network between branches and the data centre with central software policies, independent of hardware. In multi-site organisations it uses several links intelligently, cutting cost and simplifying management.

IT governance and regulation

COBIT

ISACA's framework for the governance and management of enterprise IT. It separates governance (deciding what to do) from management (building and running it); the current version is COBIT 2019.

EU Artificial Intelligence Act

The European Union regulation that governs AI systems by risk level (prohibited, high-risk, transparency obligations, minimal risk). It can also apply to non-EU companies that place products or services on the EU market.

ITIL

A best-practice framework for IT service management. Through processes such as incident, problem, change and service-level management it aims for IT to deliver reliable services to the business.

KVKK (Turkish data protection law)

Türkiye's Law No. 6698 on the Protection of Personal Data, broadly comparable to the GDPR. For AI projects working with personal data such as camera footage it imposes notice, consent, retention and security obligations.

Infrastructure and web

Core Web Vitals

Google's three core metrics of real user experience: loading speed (LCP), responsiveness to interaction (INP) and visual stability (CLS). They are used as a page-experience signal in search ranking.

Data deduplication

Storing identical data blocks only once in storage and backup. A ratio such as 2.9:1 means the logical data occupies roughly a third of the physical space.

Hyperconverged infrastructure (HCI)

A data-centre architecture in which compute, storage and networking are combined under one software layer and run on standard servers. Compared with separate SANs and storage arrays it simplifies deployment and growth.

LCP (Largest Contentful Paint)

The time until the largest image or text block on the page is rendered. Under 2.5 seconds is considered good; it is usually determined by the size of the main image and how late it is discovered.

Writing →Achievements →About →

AI and language models

Computer vision

IoT and networking

IT governance and regulation

Infrastructure and web

Let's talk.