AI project terms in plain words

A large language model that runs on servers the company operates itself, in its own data centre, so prompts, documents and answers do not leave its network. It usually means an open-weight model served on the company's own GPUs.

On-premise LLMs for companies

A language model whose trained weights are published, so a company can download it and run it on its own hardware. Licences differ, so commercial use has to be checked for each model.

On-premise AI

A service that sits between applications and language models and gives them one API. It holds an API key and a budget for each team, routes requests to the allowed models, switches to a fallback when a provider fails, and logs usage and cost. Also called an AI gateway.

What an LLM gateway does

An open-source LLM gateway and Python library that offers one OpenAI-compatible API in front of many model providers and self-hosted models, with virtual keys, budgets, routing and spend tracking. Some features need a paid enterprise licence.

LiteLLM consulting

An open-source engine for serving language models on GPUs to many users at once. It batches concurrent requests and manages GPU memory so that one server answers far more requests than a single-user tool.

vLLM on Kubernetes

A way of answering questions from your own documents: the system first searches an index for the passages that hold the answer, then gives those passages to the model, which answers from them and can cite them.

RAG development

A program in which a language model decides the next step itself, such as which tool to call or which document to read, until a task is done. In companies, agents usually prepare work and hand decisions to a person.

AI agents and process automation

A system built from one or more agents together with the tools, data access, permissions, approvals and tests around them. The model is one part. Most of the engineering goes into everything around it.

AI agents and process automation

Everything around a coding agent or other agent that shapes its work: instructions, allowed commands, permissions, checks it must pass and the review rules. Two teams with the same model get different results mainly because of their harness.

Harness engineering

A repeatable test of an AI system on a fixed set of cases that shows how often the system gets it right. Answers are checked in code or by a judge model, another LLM that grades them. Evals run before every change to a model, prompt or data source, like regression tests for normal software.

AI evals engineering

A test before go-live against criteria agreed in advance, such as a required pass rate for each kind of answer, measured on 100 to 200 cases from the company's own work, with a confidence interval around each result.

AI system testing

A service that offers tools to AI assistants and agents through the Model Context Protocol, for example to read an order from the ERP or create a ticket, with its own permissions and logs.

MCP server development

The operational practice for systems built on language models: versioned prompts and configuration, staged model updates with regression tests, monitoring of answer times, errors, quality and cost, and runbooks for incidents.

LLM platform and MLOps

Further training of an existing model on a company's own examples, so it learns a format or a narrow task. For keeping facts current, retrieval (RAG) is the better tool.

Custom model training

An agent that writes and changes code in a repository, runs commands and tests, and proposes changes for review, such as Claude Code, OpenAI Codex, Cursor or GitHub Copilot in agent mode.

Coding agent adoption