AWS AI/ML Landscape and Workflow

Playback speed

Swipe, scroll or use ← →

Cloud computing, in one idea

Every model in this course has to run somewhere, and that somewhere decides how much a team can afford to try. Cloud computing delivers IT resources over the internet on demand with pay-as-you-go pricing. Instead of buying and maintaining physical data centers and servers, an organization can obtain resources such as compute, storage, and databases when needed.

The contrast is easiest to see side by side.

Question	Owning the hardware	Using cloud resources
Where does capacity come from?	Bought and installed before the work starts	Requested at the moment it is needed
What is paid for?	The whole installation, all the time	The resources actually used
Who runs the servers?	The organization	The cloud provider

1 / 23

Cloud computing, in one idea

Machine learning is where that difference bites hardest. Training can demand a large cluster for a few hours and nothing at all for the rest of the week. AWS cloud services can scale capacity up or down, and customers pay for what they use while retaining control of their computing resources. Buying for the busiest hour means paying for that hour during every quiet one.

AWS combines cloud infrastructure with cloud and AI capabilities, including a managed machine learning suite. The same account that supplies storage and compute also supplies the AI services this course covers: nothing extra to acquire, and no hardware to install before the first experiment.

2 / 23
Quick check

A retailer trains a model on a large cluster for two days a month and needs almost nothing in between. Which cloud characteristic answers that pattern?
1. ABuying enough hardware in advance to cover the busiest possible day
  
  Buying for the peak keeps the cost of every quiet week, which is the expense on-demand resources remove.
2. BMoving the workload into a private server room the retailer maintains
  
  A private server room hands the maintenance and the fixed capacity back to the customer.
3. CScaling capacity in either direction and paying for what is used
  
  Right. AWS cloud services scale capacity up or down, and customers pay for what they use while keeping control of their computing resources.
3 / 23

Two starting points on AWS

Two services frame everything that follows, and it pays to keep their roles apart from day one.

Amazon SageMaker AI supports the tasks in a typical machine learning workflow. That workflow creates example data, trains and evaluates a model, deploys it, and continues monitoring it.

Amazon Bedrock is a fully managed service that provides access to foundation models for building and scaling generative AI applications.

Service	What it is for	What the team brings
SageMaker AI	Building a model through the workflow: data, training, evaluation, deployment, monitoring	Example data and the decisions at each stage
Amazon Bedrock	Building on foundation models that already exist	The application and the way it uses the model

4 / 23

Quick check

A team is starting a generative AI application. What does Amazon Bedrock give them?
1. AManaged access to foundation models
  
  Right. Bedrock is a fully managed service that provides access to foundation models for building and scaling generative AI applications.
2. BA cleaned training set that removes the need to choose anything
  
  Bedrock supplies model access, not a dataset, and the application decisions still belong to the team.
3. CGraphics hardware the team installs in its own server room
  
  A fully managed service is the opposite of hardware the customer has to install and operate.
5 / 23
Two starting points on AWS

Neither service replaces the other, and neither is a shortcut past the thinking.

SageMaker AI gives the workflow context used in this first part of the course, while Bedrock introduces the foundation-model path that appears later. A team that must train a model on its own examples follows the workflow. A team that wants to build on models that already exist, without operating the infrastructure behind them, starts from managed foundation-model access.

It also helps to keep the workflow stages out of this decision. Cleaning data prepares examples for training, and monitoring watches a model that is already running. Useful as both are, neither one is a way to obtain a model in the first place.

6 / 23
Quick check

A team must ship a generative AI application, wants models from several providers, and does not want to operate the model infrastructure. Where does it start?
1. AWith data cleaning, because prepared examples stand in for model access
  
  Cleaning is a preparation step for training examples and gives the team no model to build on.
2. BWith Amazon Bedrock, because it provides managed foundation-model access
  
  Right. The constraints name a managed path to existing foundation models, which is what Bedrock provides.
3. CWith monitoring, because behaviour is checked before choosing a model
  
  Monitoring belongs to a model already in production, so it cannot be the starting point.
7 / 23
Keep your progress in the app

That’s 3 of 10 quick checks. In the app they stay answered, and every lesson remembers where you left off.

Get the app Keep reading here
The workflow at a glance

The path to a model is usually drawn as three stages in a circle: generate example data, train a model, deploy the model. The circle matters as much as the stages, because the last one feeds back into the first.
1. Generate example data — decide what the model must predict, then gather and prepare the examples.
2. Train a model — train it, then evaluate whether it is good enough.
3. Deploy the model — put it where an application can request inferences.
4. Monitor and improve — watch the results, collect better data, and retrain.
A typical machine learning workflow begins by generating example data. Keep this map at hand when later lessons introduce algorithms, models, training, inference modes, and foundation models: every one of them lands on one of these stages.
8 / 23
Quick check

Which task opens the typical machine learning workflow?
1. ADeploying a model so that an application can call it
  
  Deployment comes after training and evaluation, so it cannot be the opening stage.
2. BGenerating the example data the model will learn from
  
  Right. The workflow begins by generating example data, and what that data contains depends on the problem.
3. CWatching a model already in production for signs of drift
  
  Drift can only be observed once a model is deployed and answering requests.
9 / 23
Prepare the example data

The first stage is not a technical decision, it is a business one. The data required for training depends on the business problem and on the inferences the model must generate. A model that reads handwritten digits from images needs example images of handwritten numbers; no other collection will teach it that task.

Data may come from internal repositories or public datasets and is typically brought into one repository. Consolidating first is what makes everything after it possible: an inconsistency is hard to notice across three systems that never meet.

10 / 23

Prepare the example data

Data preparation includes fetching the data, cleaning inconsistent values, and transforming attributes when a different representation can improve model performance.

Step	What happens	Example
Fetch	Pull the internal or public datasets into a single repository	Two regional sales exports land in the same place
Clean	Inspect the values and make them consistent	A country stored as both `United States` and `US` becomes one value
Transform	Change the representation so the model learns better	Temperature and humidity are combined into one attribute for a model that predicts when an aircraft needs de-icing

Cleaning and transforming are different jobs. Cleaning repairs values that should already have agreed; transforming builds a new representation that no single original column carried.

11 / 23

Prepare the example data

Order matters here. This stage comes before model training because the algorithm learns from the examples it receives, and an inconsistency left in the data is simply another pattern the model will learn.

So the sequence runs fetch, then clean, then transform: consolidate the examples, make the values agree, and only then build the representation the model will train on.

12 / 23
Quick check

A team has examples in three repositories, the same country written two ways, and two columns that work better combined. What order prepares the data?
1. ATransform the two columns, then deploy, then fix the country values later
  
  Transforming before consolidating and cleaning builds a new representation on top of values that still disagree.
2. BDeploy an untrained model first so that production traffic reveals the problems
  
  Nothing can be deployed before a model exists, and production traffic does not repair source examples.
3. CFetch everything into one repository, clean the values, then transform the columns
  
  Right. Preparation fetches the datasets into one repository, cleans inconsistent values, and then transforms attributes for better performance.
13 / 23

Train and evaluate

Training requires an algorithm or a pretrained base model, and the choice depends on the problem. Some problems are served by a built-in algorithm; others start from a model that has already been trained on a broad corpus and is adapted from there.

Training also requires compute resources, whose scale depends on the dataset size and how quickly results are needed — anything from a single general-purpose instance to a distributed cluster.

Input	Why training needs it
Prepared examples	The material the model actually learns from
An algorithm or a pretrained base model	The method that turns those examples into a model
Compute resources	The capacity that runs the job, sized by data volume and deadline

14 / 23

Quick check

Besides prepared examples, what does a training run need?
1. AAn algorithm or a pretrained base model, plus compute resources
  
  Right. Training needs a method — an algorithm or a pretrained base model — and compute sized to the dataset and the deadline.
2. BA monitoring report covering production traffic the model has not yet served
  
  Monitoring reports describe a model that is already answering requests, so they cannot feed a first training run.
3. CAn application that has already been deployed to its end users
  
  Deployment follows training; an application cannot be the input that produces the model.
15 / 23
Train and evaluate

Evaluation follows training and determines whether the accuracy of the model's inferences is acceptable. A model that does not meet the required quality should not move forward merely because training completed.

Treat evaluation as a gate rather than a formality. When the accuracy is not acceptable, the fix almost always sits in the earlier stages: more or better examples, a different algorithm or base model, or a different representation of the attributes. Deploying anyway does not improve the answers; it only puts them in front of users.

16 / 23
Quick check

A training job has just finished. What does the evaluation step decide?
1. AWhether monitoring can be dropped from the cycle for good
  
  Monitoring stays in the cycle after deployment; evaluation does not remove it.
2. BWhether the organization should buy its own data center
  
  Owning infrastructure is a cloud decision and has nothing to do with the quality of a model's answers.
3. CWhether the accuracy of the model's inferences is acceptable
  
  Right. Evaluation follows training and asks whether the accuracy of the inferences is good enough to go on.
17 / 23
Deploy the model

Deployment integrates the trained model with an application so the application can request inferences. This is the moment a model stops being the output of a project and starts answering new requests and records.

A model was traditionally re-engineered before being integrated with the application code. SageMaker AI hosting can deploy a model independently from application code, which decouples the model from that code. Decoupled, the model can be replaced on its own schedule while the application keeps calling it in exactly the same way.

18 / 23
Quick check

What does deployment change, and what does SageMaker AI hosting make easier about it?
1. AIt cleans whatever inconsistent values were left behind in the prepared training examples
  
  Cleaning happens while preparing examples, long before a model is integrated with an application.
2. BThe application can request inferences, and hosting deploys the model separately from application code
  
  Right. Deployment integrates the model so the application can request inferences, and SageMaker AI hosting decouples the model from that application code.
3. CIt replaces evaluation entirely, on the grounds that a deployed model has already proved its accuracy
  
  Evaluation happens before deployment, and deploying a model proves nothing about its accuracy.
19 / 23
Keep the cycle running

Deployment is not the end of the workflow. After deployment, teams monitor inferences, collect more high-quality data, and evaluate the model to identify drift. They can then update the training data and retrain the model as more examples become available.

Machine learning is a continuous cycle rather than a one-time sequence that stops at deployment. The world the model reads keeps moving, so a model that was accurate at launch can become less so without anything breaking. The loop closes where it opened: today's real inputs become tomorrow's training examples.

20 / 23
Quick check

A deployed model is answering less accurately, and the team has since collected better examples. What continues the cycle?
1. AEvaluate for drift, update the data, and retrain
  
  Right. Monitoring and drift evaluation are what tell the team when new high-quality data should feed a retraining run.
2. BRedeploy the same model unchanged and skip the evaluation this time
  
  Redeploying an unchanged model repeats the same answers, and skipping evaluation removes the only quality gate.
3. CLeave the model as it is and stop monitoring its inferences
  
  Turning off monitoring discards the very signal that showed the accuracy had slipped.
21 / 23
Key takeaways
- AWS provides on-demand cloud resources and AI capabilities without requiring customers to own the underlying physical infrastructure.
- SageMaker AI supports a typical model workflow, while Amazon Bedrock provides managed access to foundation models for generative AI applications.
- The model workflow prepares example data, trains and evaluates a model, deploys it, and then monitors and retrains it as needed.
- Deployment begins model use; monitoring and new data continue the machine learning cycle.
Read the stages as one sentence: a business problem decides the data, the data feeds training, evaluation decides whether the model may go on, deployment puts it to work, and monitoring sends it back to the start.
22 / 23
Quick check

Which summary puts the roles in the right order?
1. ABedrock cleans example data, evaluation deploys the model, and monitoring ends the workflow
  
  Bedrock provides foundation-model access rather than data preparation, and monitoring continues the cycle instead of closing it.
2. BDeployment selects the algorithm, cleaning measures accuracy, and cloud capacity is fixed once bought
  
  The algorithm is chosen for training, accuracy is measured by evaluation, and cloud capacity can be scaled in either direction.
3. CPrepared data feeds training, evaluation gates deployment, and monitoring feeds retraining
  
  Right. That is the continuous cycle: data, training, evaluation as the gate, deployment, and monitoring that returns new examples to training.
23 / 23
10 quick checks · then the test

In the app, finishing the quick checks opens this lesson’s 10-question test, and the ones you miss come back exactly when you’re about to forget them.

Take this lesson’s test → Next lesson: AIF-C01 Exam Overview Get the app

← Previous Keep your progress · Get the app Next card →

The whole course, on your phone

Lessons you can read, audio you can listen to on the way to work, and practice that remembers what you got wrong.

Get the app

AWS AI/ML Landscape and Workflow

Cloud computing, in one idea

Two starting points on AWS

Keep your progress in the app

The workflow at a glance

Prepare the example data

Train and evaluate

Deploy the model

Keep the cycle running

Key takeaways

10 quick checks · then the test

The whole course, on your phone