aws AI Practitioner services & lifecycle

AWS Region

A physical location in the world where AWS clusters its data centres.

• Each Region has ≥ 3 Availability Zones (the magic number — not 2)
• AZs are isolated, physically separate, geographically grouped
• Connected by redundant, ultra-low-latency networks

Pick a Region for: compliance (data residency), latency (close to users), service availability, price.

Availability Zone (AZ)

One or more discrete data centres with redundant power, networking and connectivity, inside a Region.

• Each AZ has independent power, cooling and physical security
• AZs in a Region are interconnected via high-bandwidth, low-latency, redundant metro fibre
• Naming: eu-west-2a, eu-west-2b, etc.

Spread workloads across ≥ 2 AZs in the same Region for high availability — single AZ = single point of failure.

Edge Locations

Separate from Regions and AZs — these are content delivery endpoints close to end users.

• Used by services like CloudFront (CDN) and Route 53
• Cache content near users for faster delivery
• Far more numerous than Regions (400+ globally)

Hierarchy: Region contains AZs which contain data centres. Edge Locations sit outside this stack.

SageMaker Canvas

Generate ML predictions without writing any code. Visual, drag-and-drop interface for business analysts.

• Chat with popular LLMs
• Use Ready-to-use models
• Build custom models on your data — automatically

Think of it as the friendly front-end to ML for people who don't live in PyCharm.

SageMaker JumpStart

An ML hub of foundation models, built-in algorithms and prebuilt solutions you can deploy in a few clicks.

• Access to popular FMs (Llama, Stable Diffusion, etc.)
• Pre-built end-to-end solutions for common use cases
• Fine-tune models on your own data

Canvas vs JumpStart: Canvas = no-code UI for predictions. JumpStart = model + solution hub for builders.

SageMaker Clarify

Two jobs, both critical for Responsible AI:

• Detects bias in your data and in model predictions across groups
• Explains predictions — shows how each input feature contributed (SHAP values)

Exam keyword: "bias", "fairness", "explain predictions", "transparency" → Clarify.

SageMaker Model Monitor

Continuously watches deployed models in production and alerts when something goes wrong.

• Detects data drift — when incoming data shifts from training data
• Detects concept drift — when the underlying relationships change
• Flags data quality issues and anomalies
• Alerts you to inaccurate predictions

Clarify catches bias upfront and explains predictions. Model Monitor catches problems after deployment. Both = Responsible AI.

SageMaker Data Wrangler

The fastest, easiest way to prep tabular and image data for ML — with little to no code.

• Import data from many sources (S3, Athena, Redshift, Snowflake)
• 300+ built-in transforms (impute, encode, normalise)
• Visualise distributions and outliers
• Export as a pipeline or feature set

Different from Ground Truth, which is for labeling data (annotating it), not transforming it.

SageMaker Ground Truth

For labeling and annotating training data — turns raw data into labeled data for supervised learning.

• Use Mechanical Turk workers, vendors, or your own team
• Active learning: ML auto-labels easy items, humans handle hard ones
• Reduces labeling cost vs. doing everything by hand
• Also: Ground Truth Plus = fully managed labeling service

Exam keyword: "data labeling", "annotate data", "human reviewers add labels" → Ground Truth.

The 6 pillars (AWS Responsible AI)

AWS tools that map here: Clarify (fairness + explainability), Model Monitor (safety + robustness), A2I (controllability).

Amazon Augmented AI (A2I)

Implements human review for ML predictions — useful when accuracy matters and the model is uncertain.

• Built-in workflows for Rekognition, Textract, custom models
• Trigger human review when prediction confidence is low
• Or sample a % of predictions for QA review
• Reviewers can be your team, vendors, or Mechanical Turk

This is the AWS service for the "human-in-the-loop" pillar of Responsible AI. Critical for high-stakes use cases like medical, legal, content moderation.

The ML process sequence

The fundamental 4-step sequence the exam will test:

1. Data collection — gather from all sources
2. Data preprocessing — clean, normalise, split train/val/test
3. Model training — feed prepared data to the algorithm
4. Model evaluation — assess on the held-out test set

In the wider ML lifecycle (MLOps), this extends to: framing → data → train → deploy → monitor → iterate.

Training

The phase where the model learns — its internal parameters (weights) are being updated.

• Uses the training set (~60–80% of data)
• Algorithm minimises an error function
• Model "sees" examples and adjusts
• The output is a trained model

Validation vs Testing

Mnemonic: validation = the dress rehearsal (you adjust). Testing = the actual show (you don't tweak after).

Neither involves serving predictions to real users — that's inference.

Inference

The deployed model uses its frozen learned parameters to make a prediction on new, real-world input.

• This is what users actually experience
• No learning happening — parameters don't change
• Two modes: real-time (live API, low latency) and batch (process many records at once)

In SageMaker, you deploy a model to an endpoint for real-time inference, or run a batch transform job for bulk.

Foundation Models use self-supervised learning

The defining technique for FMs. Model gets vast amounts of unlabeled data, then generates labels from the data itself.

• No humans labeling anything upfront
• Example: LLM predicts the next word — the "label" is just the next word that already exists in the text
• Used to pre-train, then fine-tune for downstream tasks

Question keywords: "foundation model" + "generates labels" + "from raw input" → always self-supervised.

The 4 learning paradigms

Plus semi-supervised = small labeled + large unlabeled, uses pseudo-labeling.

Amazon Bedrock

The fully managed service for accessing foundation models via API from multiple providers.

• Access to Anthropic Claude, Meta Llama, Mistral, Cohere, Amazon Titan, etc.
• Serverless — no infrastructure to manage
• Fine-tune FMs on your own data privately
• Knowledge Bases for RAG, Agents for tool-use
• Guardrails for safety filtering

Bias & Variance, locked in

Memorable framing:

• Underfit = high bias, model is too simple, biased toward being basic
• Overfit = high variance, model memorised noise, output swings wildly with new data

Goldilocks zone in the middle. Reduce overfit with: more data, regularisation, early stopping, dropout, ensembling.

The biggest ML challenge: data

Collecting and preparing high-quality data is THE primary challenge in real-world ML implementation.

• Algorithms are commodity (open-source, well-documented)
• Compute is rentable (AWS, GPUs on demand)
• Data quality, quantity, labeling, bias — all bottlenecks
• "Garbage in, garbage out" — your model is only as good as its data

Adjacent challenges: model interpretability, deployment costs, monitoring drift, ethical concerns, regulatory compliance.