Developing ML Solutions with Amazon SageMaker AI
One workshop for the whole job
The story: Instead of renting a separate place to store wood, a place to cut it, a place to assemble furniture, and a showroom, you rent one big workshop that has all of them under one roof, with one front desk.
In AI/AWS terms: Amazon SageMaker AI is that workshop: a fully managed ML service for the whole workflow in one interface, from collecting and preparing data to building, training, deploying, and monitoring. SageMaker Studio is the front desk: the recommended web-based UI for all of it.
For the exam: SageMaker AI covers the whole ML workflow. SageMaker Studio is its web-based UI.
The stations in the workshop
The story: Walk through the workshop in order:
- People sticking name tags on raw materials.
- A prep bench where wood is cleaned and cut, with simple point-and-click tools.
- A shared shelf of pre-cut parts every carpenter can reuse.
- An inspector checking for unfairness and explaining why each piece looks the way it does.
- A drag-and-drop design corner for people who've never used a saw.
- A catalog of ready-made furniture you can adjust.
- A robot that tries many designs automatically and picks the best.
- The main building floor, with a helper that fine-tunes the machine settings.
- A notebook logging every attempt.
- A registry of approved designs, each with a version number.
- A conveyor belt connecting every step automatically.
- A quality inspector watching what leaves the shop.
In AI/AWS terms:
| Station | Stage | Feature |
|---|---|---|
| Name tags | Label data | SageMaker Ground Truth: human labeling, including RLHF |
| Prep bench | Prepare data | SageMaker Data Wrangler: low-code import, cleaning, transformation, and analysis |
| Shared shelf | Store features | SageMaker Feature Store: shared, reusable features for training and inference |
| Fairness inspector | Detect bias and explain | SageMaker Clarify |
| Drag-and-drop corner | Build without code | SageMaker Canvas: no-code visual ML for business analysts |
| Ready-made catalog | Start from pre-trained models | SageMaker JumpStart |
| Design robot | Automate model building | SageMaker Autopilot: AutoML |
| Building floor | Train and tune | Managed training jobs and automatic model tuning (hyperparameter optimization) |
| Logbook | Track experiments | SageMaker Experiments |
| Design registry | Version and approve models | SageMaker Model Registry |
| Conveyor belt | Orchestrate pipelines | SageMaker Pipelines: CI/CD for ML |
| Quality inspector | Monitor in production | SageMaker Model Monitor |
For the exam: No-code for business analysts = Canvas. AutoML = Autopilot. Labeling = Ground Truth. Reusable features = Feature Store. Versioning and approval = Model Registry. CI/CD = Pipelines.