Every word your vendor uses, in plain English.

AI

Agent

A model given tools and a goal, allowed to decide the steps itself rather than following a fixed script.

Delivery

AI Blueprint

The document that fixes architecture, model choice, hosting and monthly cost before any code is written.

Security

Air-gapped

A system with no network path to the outside world. Updates arrive by hand, on purpose.

Infrastructure

API Gateway

The single front door that routes, authenticates and rate-limits every call into your services.

Infrastructure

Autoscaling

Capacity that grows with load and, more importantly, shrinks again when the load goes away.

Security

Backup & DR

Copies of your data plus a tested plan to run the business from them. Untested backups are wishes.

Infrastructure

Bare metal

A physical server dedicated to you, with no hypervisor between your workload and the hardware.

AI

Benchmark

A fixed set of tasks used to compare models. Useful for shortlisting, misleading as a purchase decision.

DevOps

Blue-green deployment

Running two identical environments and switching traffic between them, so a rollback is a routing change.

AI

Chunking

Splitting documents into passages small enough to retrieve precisely and large enough to still make sense.

DevOps

CI/CD

Automated build, test and release. The reason a fix can ship on Tuesday instead of next quarter.

Infrastructure

Container

An application packaged with everything it needs to run, so it behaves the same on every machine.

AI

Context window

How much text a model can consider at once. Everything outside it may as well not exist.

Data

Data lineage

A record of where each piece of data came from and what transformed it on the way.

Compliance

Data residency

A requirement that data physically stays inside a country or region. It shapes architecture more than any model choice.

Security

Disaster recovery

How fast you can be running again after a serious failure, and how much data you accept losing.

AI

Drift

When the real world moves away from the data a model was trained on and accuracy quietly decays.

Infrastructure

Edge inference

Running a model on hardware near the user or machine, so it keeps working when the network does not.

AI

Embedding

A numeric fingerprint of a piece of text or an image, used to find things by meaning rather than keyword.

AI

Evaluation harness

A regression suite for model behaviour. It turns a model upgrade from a gamble into a reviewable diff.

AI

Fine-tuning

Further training an existing model on your own examples so it adopts a style, format or narrow skill.

Infrastructure

GPU

The processor that makes training and inference practical. Usually the largest single line on an AI bill.

AI

Guardrails

Checks around a model that block, redact or escalate outputs before they reach a person.

AI

Hallucination

A confident, fluent answer that is simply untrue. The reason grounding and citations matter.

Infrastructure

High availability

Designing so that a single failure does not take the service down. Costs more; sometimes worth it.

Infrastructure

Hybrid cloud

Running some workloads in a public cloud and some in your own datacentre, on purpose rather than by accident.

AI

Inference

Actually using a trained model to produce an answer. The recurring cost, as opposed to training's one-off cost.

DevOps

Infrastructure as code

Defining servers and networks in files you can review and version, instead of clicking through a console.

Infrastructure

Kubernetes

A system for running containers across many machines. Powerful, and heavier than most teams first expect.

Infrastructure

Latency

How long a request takes end to end. Users feel this long before they notice accuracy.

AI

LLM

A large language model: a system trained on text that predicts what should come next, well enough to be useful.

DevOps

MLOps

The practice of getting models into production and keeping them healthy there, not just training them.

Architecture

Multi-tenancy

One system serving many customers with strict boundaries so no one ever sees anyone else's data.

DevOps

Observability

Enough logs, metrics and traces to answer questions you did not anticipate when the system broke.

Infrastructure

On-premise

Hardware in your own building or your own rack. Still the right answer for some regulated workloads.

AI

Open-weight model

A model whose parameters you can download and run yourself, on infrastructure you control.

Security

Penetration test

Paying someone to attack your system on purpose, before somebody does it without an invoice.

Delivery

POC

A proof of concept: the smallest build that answers whether an idea works, before anyone funds the real thing.

Infrastructure

Private cloud

Cloud-style capacity dedicated to one organisation, rather than shared with everyone else on the platform.

Security

Prompt injection

An attack where hostile text inside a document persuades a model to ignore its actual instructions.

AI

Quantisation

Shrinking a model's numeric precision so it runs on cheaper hardware, trading a little accuracy for a lot of cost.

AI

RAG

Retrieval-Augmented Generation: fetching your own documents and giving them to the model, so answers cite real sources.

Read the full note →

Infrastructure

Rate limiting

Capping how often a caller can hit your service, so one client cannot degrade it for everyone.

Security

Red teaming

Deliberately trying to make an AI system misbehave, so you find the failure before a customer does.

Delivery

SLA

A written commitment to availability and response times, with consequences attached. Without consequences it is a brochure.

AI

Token

The unit models read and bill in — roughly three-quarters of a word. Cost and context limits are both counted in these.

Data

Vector database

Storage built to find things by meaning, using embeddings, rather than by exact keyword match.

Security

Zero trust

Assuming the network is already hostile, so every request is authenticated rather than trusted by location.