Course Introduction And Orientation

Main Point

Trace the meaning, movement and permitted use of information so an AI application can use the right evidence for its task. The course project joins those decisions in one data architecture.

Key Points

• Explain what a record can establish before asking a model to use it.
• Follow records through storage, transformation and fitness checks.
• Separate training, evaluation, runtime evidence and application state.
• Design updates, permissions and accountability into the same architecture.

1. Why this course begins with records

An AI system does not receive a business, customer, or transaction directly. It receives representations selected by people and software: fields, rows, documents, images, events, labels, retrieval results, and instructions. The choices behind those representations affect what the system can answer and what its answers can support.

This course follows one recurring example: a company wants an assistant to answer questions about orders and returns. A record says an order was shipped. A customer says it never arrived. A scan event may provide evidence about a particular scan, but it does not automatically establish the full delivery outcome. A return policy is a rule that applies under defined conditions; it is not another observation of the shipment. Keeping these roles distinct prevents the system from treating unlike evidence as interchangeable.

2. The course route

The first section asks what a record means and how records are connected. You will examine fields, values, units, timestamps, table grain, keys, joins, and missing values. These details determine what a query returns and whether a number or status can be interpreted.

The second section follows information through operational systems and analytical stores. You will compare storage choices, ingestion and transformation patterns, pipeline layers, and lineage. A value can change as it moves; an answer needs a trace back to the source and processing steps that produced it.

The third section asks whether data is fit for a specific use. Quality is not one score that applies everywhere. Sampling, representation, labels, training examples, evaluation design, synthetic data, and proprietary data claims each need a stated purpose and limits.

The final section distinguishes four roles. Training data is used in a learning process that changes model parameters. Evaluation data measures behavior on cases reserved for that purpose; development validation and a final held-out test should not be casually treated as the same set. Retrieval supplies selected evidence at request time. Application state retains scoped information such as a task status or user preference. These mechanisms differ in freshness, update, access, and deletion behavior.

3. Your project

Choose the orders-and-returns assistant or another narrow business use. Map the information it needs, where that information originates, how it changes, and who may access it. Separate training, evaluation, retrieval, and state. Add checks for missing, stale, duplicate, incorrect, and unauthorized information. Then explain how the system should respond when sources conflict or evidence is absent.

You do not need to code. Small examples will let you reason about what one record represents, which convention gives a field meaning, and what evidence supports a decision. The goal is a design that makes information understandable, current enough, authorized, and testable for its particular task.

4. Bringing the lesson together

The opening shipment question has several evidence paths: a recorded status, a scan and a policy each support a different part of the answer. Your architecture will make those paths visible, keep their limits clear and show who maintains them. Start by defining what one record represents.

Learning Checks

Attempt each answer before expanding the reasoning.

M13-INTRO-Q01

The four items are:

• Examples used to update model parameters.
• Reviewed final test cases held for measuring the frozen candidate.
• The approved policy version and order facts selected for the current customer question.
• A scoped preference for how a customer wants an explanation.

Classify the four listed items as training, evaluation, request-time retrieval, or application state. Which provide current evidence for the order question, and which should never be the default production answer source?

M13-INTRO-Q02

After deleting an order row from its original database, what other copies or derived representations should the team check?

Optional background

AI Foundations introduces data and instructions. How Machine Learning Learns develops training and evaluation. MLOps And The AI Development Lifecycle is planned background for pipelines and versioning; its lessons are not yet available. These are optional connections, not prerequisites for this course.

Sources

• HF-TRAIN Hugging Face. *Fine-tuning*. https://huggingface.co/docs/transformers/v4.57.0/en/training. Transformers 4.57.0. Retrieved 29 October 2026.
• SK-LEAK scikit-learn developers. *Common pitfalls and recommended practices*. https://scikit-learn.org/1.7/common_pitfalls.html. scikit-learn 1.7.2. Retrieved 29 October 2026.
• AZ-RAG Microsoft. *Retrieval-augmented generation in Azure AI Search*. https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview. Live architecture documentation; revision date not asserted. Retrieved 29 October 2026.
• LG-MEM LangChain. *Memory overview*. https://docs.langchain.com/oss/python/concepts/memory. Live framework documentation; version/date not asserted. Retrieved 29 October 2026.

The business scenarios and numerical counts are hypothetical teaching examples. These public references support the underlying methods; they do not report results for the fictional service.