The words, defined once.

Agentic AI is a system in which a model plans a task, calls tools and other systems, checks its own work and finishes a multi-step job, instead of answering a single prompt.

At Zingaro AI: Zingaro AI builds agents with tools, permissions, confidence thresholds and a review queue, so uncertain items go to a person. Agentic AI development

AI-native services are work delivered as an outcome by a company whose workforce is models and agents, with people on oversight, rather than software sold for the client's own staff to operate.

At Zingaro AI: It is the category Zingaro AI belongs to: the client pays for the job done, in its own systems. What AI-native services means

An air-gapped deployment runs on machines with no network route to the internet at all, so models, packages and updates have to be carried in as verified bundles.

At Zingaro AI: Zingaro AI packages, signs and rolls model updates through air-gapped networks in stages, with the previous version kept for rollback. Updating models inside an air-gapped network

Automatic speech recognition is the conversion of spoken audio into text by a model; its accuracy is usually measured as word error rate, the share of words it gets wrong.

At Zingaro AI: Zingaro AI trains speech recognition on the client's own line quality and dialect, and measures it on the client's own recordings. Custom ASR and TTS training

Barge-in is a caller speaking while the system is still talking; a good voice agent stops within a fraction of a second and keeps listening without losing the caller's words.

At Zingaro AI: Zingaro AI treats barge-in as a first-class event, with playback cancelled and the recogniser kept running throughout.

A computer-use agent is an AI agent that operates software through its screen, reading the interface and clicking and typing as a person would, for systems that have no API.

At Zingaro AI: Zingaro AI uses computer-use agents for legacy systems, with strict permissions and every action logged.

Distillation is training a smaller model to reproduce the outputs of a larger one on a chosen task, so the small model keeps most of the large model's skill on that task at a fraction of the cost.

At Zingaro AI: Zingaro AI distils large models into small ones for production and routes the uncertain cases to a larger model or a person. The case for small models

Endpointing is a voice system's decision that the speaker has finished talking; too eager and the agent interrupts, too patient and every turn carries a dead pause.

At Zingaro AI: Zingaro AI combines short acoustic silence with whether the partial transcript looks complete, which is how people do it.

An evaluation suite is a fixed set of test cases with expected outcomes, run against an AI system before every release to measure whether a change made it better or worse across the whole distribution of inputs.

At Zingaro AI: Zingaro AI gates every release on its evaluation suite; a rising training curve is not evidence, the suite is. Evals as a release gate

Fine-tuning is continuing the training of an existing model on a smaller, task-specific dataset so that its weights change to fit that task, format or style.

At Zingaro AI: Zingaro AI fine-tunes on the client's own examples, then gates the result with an evaluation suite before it ships. Fine-tuning vs RAG vs prompting

Guardrails are the checks around a model that stop unsafe, off-policy or malformed outputs from reaching a user or a system: input filters, output validators, permission limits and hand-off rules.

At Zingaro AI: Zingaro AI builds guardrails as code and tests them in the same suite as everything else. Evals, red-teaming and guardrails

Human in the loop is a design in which a person reviews or decides part of an automated process, typically the items the system is unsure about, so that mistakes are caught before they cost money or trust.

At Zingaro AI: At Zingaro AI the loop is a queue: agents take the volume, reviewers hold the exceptions, and the record shows which did what. Agents need a queue, not a chat window

Inference is running a trained model to produce an output for a new input; it is the moment a client's data meets the model, and the cost that repeats on every request.

At Zingaro AI: Zingaro AI engineers inference on the client's own hardware or cloud, measured against the client's real traffic. Inference engineering

A large language model is a neural network trained on very large amounts of text to predict and generate language, which lets it answer questions, write, summarise, extract and follow instructions.

At Zingaro AI: Zingaro AI uses hosted frontier models where breadth matters and fine-tunes open-weight models where privacy, cost, latency or a narrow domain wins. Custom LLM training

The Model Context Protocol is an open standard for connecting AI agents to tools and data sources through a common interface, so a tool built once can be used by many agents.

At Zingaro AI: Zingaro AI gives its agents tool access through MCP servers with explicit permissions per tool.

A multimodal model accepts more than one kind of input, such as text and images or audio and video, and can answer questions about them together.

At Zingaro AI: Zingaro AI uses vision-language models for inspection, reading documents and events from video, on servers or edge devices. Computer vision and multimodal AI

Post-training is every training stage after a model's initial pre-training: supervised fine-tuning, preference tuning and reinforcement learning, used to shape behaviour rather than add knowledge.

At Zingaro AI: Zingaro AI uses reinforcement post-training with verifiable rewards for tasks a checker can score, such as format compliance and tool calls. RL with verifiable rewards, on one node

Quantisation is storing a model's weights and activations in lower-precision numbers, such as 8-bit or 4-bit instead of 16-bit, so the model needs less memory and runs faster with a small, measurable loss in accuracy.

At Zingaro AI: Zingaro AI quantises production models only where the evaluation suite shows the loss is acceptable for the job.

Reinforcement learning with verifiable rewards trains a model against a reward computed by code, such as a test passing or a schema validating, instead of a reward model learned from human preferences.

At Zingaro AI: It is how Zingaro AI teaches small models to follow a schema, call the right tool and hand off when unsure.

Retrieval-augmented generation is the pattern of fetching relevant documents from a store at answer time and giving them to a language model as context, so the answer is grounded in sources that can be cited.

At Zingaro AI: Zingaro AI builds retrieval over the client's own documents with cited answers and a fallback to a person. Knowledge systems and enterprise search

Sovereign AI is an AI deployment in which the model weights, the inference, the logs and the encryption keys all stay under the client's own control and jurisdiction, and updates arrive in a form the client can verify.

At Zingaro AI: Zingaro AI deploys on the client's servers or in the client's cloud account, and treats it as a normal deployment. Sovereign AI, in plain words

Speculative decoding is an inference technique in which a cheap draft proposes several next tokens and the main model verifies them in one pass, producing identical output faster when the draft is usually right.

At Zingaro AI: Zingaro AI switches it on per workload after measuring acceptance rate and throughput, because it can also slow a busy server down. Speculative decoding in production

A speech-to-speech model takes audio in and produces audio out in one model, without a separate transcription, reasoning and synthesis stage, which makes turns faster but leaves less room to insert tools and rules.

At Zingaro AI: Zingaro AI uses speech-to-speech where the conversation is simple and a streaming pipeline where the call has to touch a system of record. A latency budget for voice agents

Text-to-speech is the generation of spoken audio from text by a model, including voices trained to sound like a particular speaker or brand.

At Zingaro AI: Zingaro AI trains text-to-speech voices per client, in the languages the client's customers speak.

A voice agent is an AI system that holds a spoken conversation on a phone line or in an app, completes bounded tasks such as booking or verification in the connected systems, and hands off to a person when it should.

At Zingaro AI: Zingaro AI builds and runs voice agents on its own numbers and trunks, with people on the exceptions. Voice agent development