Plain definitions of the terms used across Epsilon's programs — what each one means, and what it is usually confused with. Written for somebody deciding whether a program is for them.
Artificial intelligence is software that performs tasks normally needing human judgment — recognizing an image, reading a document, forecasting a number, holding a conversation.
It is an umbrella term, not a single technology. Most working AI systems today are machine-learning models trained on examples, wrapped in ordinary software that decides when to call them and what to do with the answer.
See also: Machine learning (ML) Generative AI
Machine learning is the practice of fitting a model to historical examples so that it can make predictions about new cases, instead of following rules a person wrote by hand.
The distinction that matters in practice: a rules engine does what it was told, and a model does what the data implied. That makes a model better at messy problems and harder to explain, which is why governance and monitoring are part of the job rather than an afterthought.
See also: Deep learning MLOps Data governance
Deep learning is machine learning with neural networks of many layers, which learn useful features from raw data — pixels, audio, text — instead of being given features an analyst designed.
It is what made image recognition, speech and modern language models work. The cost is data and compute: a deep model usually needs far more examples than a classical one to beat it.
See also: Neural network Computer vision Natural language processing (NLP)
A neural network is a model built from layers of simple numeric units, whose connection weights are adjusted during training until its outputs match the examples it was shown.
The name borrows from biology and the resemblance stops early — it is a differentiable function with a great many parameters, fitted by gradient descent.
See also: Deep learning
A large language model is a neural network trained on very large amounts of text to predict what text comes next, which turns out to be enough to summarize, translate, classify, answer and write.
It has no database inside it and no notion of truth: it produces the most plausible continuation given its training and the prompt. That single fact explains both why it is useful and why it hallucinates.
See also: Generative AI Hallucination Retrieval-augmented generation (RAG) Fine-tuning
Generative AI is the class of models that produce new content — text, images, audio, code — rather than returning a label, a score or a number.
It is a different shape of problem from prediction, and it is evaluated differently: there is rarely one correct output, so quality is judged against criteria a person sets rather than against a held-out answer key.
See also: Large language model (LLM) Prompt engineering AI agent
Retrieval-augmented generation is the pattern of searching your own documents for the passages relevant to a question, then giving those passages to a language model and asking it to answer from them.
It is how an assistant answers about material the model was never trained on — a policy manual, a contract set, a product catalog — and why it can cite a source. It reduces hallucination but does not remove it: a model given the wrong passage will answer confidently from the wrong passage.
See also: Large language model (LLM) Embedding Vector database Hallucination
Prompt engineering is writing the instructions given to a language model — the task, the context, the format, the constraints — so that its output is useful and repeatable.
In production it is less about clever phrasing than about being explicit: stating the role, supplying examples, defining the output structure, and saying what to do when the answer is not in the material.
See also: Large language model (LLM) Generative AI
A hallucination is a fluent, confident statement from a language model that is not true — an invented citation, a plausible figure, a function that does not exist.
It is not a bug to be patched but a consequence of how the model works: it predicts plausible text, and plausible is not the same as correct. The practical defenses are retrieval, citation, and a human check on anything that matters.
See also: Large language model (LLM) Retrieval-augmented generation (RAG)
An embedding is a list of numbers representing a piece of text, an image or an item, positioned so that things with similar meaning sit close together.
It is what lets a search find "annual leave" when the document says "vacation entitlement" — the match is on meaning rather than on words.
See also: Vector database Retrieval-augmented generation (RAG)
A vector database stores embeddings and answers the question "which stored items are closest in meaning to this one?" quickly, across millions of records.
It is the retrieval half of a RAG system. A conventional database can do this at small scale; the specialized ones exist for the speed and the index structures, not for a different idea.
Fine-tuning is continuing a trained model's training on your own examples, so that it adopts a style, a format or a task it did not handle well out of the box.
It teaches behavior, not facts. When the goal is for a model to know your documents, retrieval is the cheaper and more maintainable answer — the documents change, and a fine-tuned model does not.
See also: Large language model (LLM) Retrieval-augmented generation (RAG)
An AI agent is a system that uses a model to decide which steps to take and which tools to call — searching, reading, writing to a system — rather than producing one answer and stopping.
The engineering question is never the model, it is the boundary: what the agent may do unattended, what it must get approved, and what it logs so somebody can check afterwards.
See also: Generative AI Robotic process automation (RPA) Data governance
Natural language processing is the field concerned with getting software to work with human language — classifying it, extracting from it, translating it, summarising it.
Arabic adds work that English does not: rich morphology, optional diacritics, and dialects that differ from the written standard — which is why a model that performs well in English can perform poorly on the same task in Arabic.
See also: Large language model (LLM) Deep learning
Computer vision is the field concerned with extracting information from images and video — what is in the frame, where it is, whether it changed, whether it is defective.
In industry it is mostly inspection, counting and safety: reading a gauge, spotting a defect on a line, noticing that somebody is in an area they should not be.
See also: Deep learning Predictive maintenance
Time-series forecasting is predicting future values of something measured repeatedly over time — demand, load, price, consumption — using its own history and anything that moves with it.
The trap is evaluating it like any other model: a time series must be tested on the future, never on a random split, or the model is scored on information it would not have had.
See also: Machine learning (ML) Predictive maintenance
Predictive maintenance is using sensor and maintenance history to estimate when a machine is likely to fail, so it is serviced before it does rather than on a fixed calendar.
It pays where unplanned downtime is expensive and where failures leave a trace in the data before they happen. Where neither holds, scheduled maintenance is the cheaper answer and saying so is part of the analysis.
See also: Time-series forecasting Computer vision
MLOps is the practice of running machine-learning models in production — versioning data and models, deploying them, monitoring what they do, and retraining them when the world moves.
It exists because a model that was accurate when it shipped quietly stops being accurate: the inputs drift, the business changes, and nothing fails loudly. Monitoring is the part most often skipped and most often regretted.
See also: Machine learning (ML) Data pipeline Data governance
Robotic process automation is software that performs a repetitive, rule-based clerical process the way a person would — opening the system, copying the field, filing the record.
It is not machine learning and does not learn; it follows a defined path exactly. Pairing it with a model — which reads the unstructured document the path depends on — is where the two meet.
See also: AI agent Machine learning (ML)
ETL is the process of taking data out of its source systems, reshaping and cleaning it, and loading it where it will be analyzed.
ELT is the same three steps in a different order — load first, transform in the warehouse — which modern cloud warehouses made practical and often cheaper.
See also: Data pipeline Big data
A data pipeline is the automated path data travels from where it is produced to where it is used, including the steps that validate, transform and schedule it.
Most of the work in an analytics or AI project is here rather than in the model, and most of the failures are too — a pipeline that silently delivers stale data produces confident wrong answers downstream.
See also: ETL (extract, transform, load) Big data Data governance
Big data describes datasets large or fast enough that ordinary single-machine tools stop working, so the processing is spread across a cluster.
The threshold is practical, not a number: it is reached when the data no longer fits the machine or the deadline, which is why the tooling matters more than the label.
See also: Data pipeline ETL (extract, transform, load)
Business intelligence is reporting and analysis of what has already happened — dashboards, metrics and the queries behind them — so that people can see the state of the business and act on it.
The line against machine learning is simple and worth keeping: BI tells you what happened and a model estimates what will. A dashboard that nobody acts on is a reporting problem, not a data problem.
See also: ETL (extract, transform, load) Machine learning (ML)
Data governance is the set of rules and records that say who may use which data for what, where each figure came from, and who is accountable for it being right.
It is what makes an AI result defensible after the fact. Without lineage and access control, a correct answer and a wrong one look identical in a slide.
See also: Data pipeline MLOps
Security Features
Verification successful
Secure · Private · Verified
