A focused, free prep hub for the AWS Certified AI Practitioner (AIF-C01). Concept walkthroughs, an interactive Bedrock + RAG architecture, 30+ Q&A, a 15-question timed mock exam, hands-on labs, real customer use cases, and full-syllabus cheat-sheet PDFs — all in one page.
Click any topic to expand. Each panel covers what it is, when AWS expects you to use it, and the exam-relevant details. Read top-to-bottom for a clean walk-through of the entire AIF-C01 syllabus.
Domain 1 is 20% of AIF-C01. AWS expects you to know the AI hierarchy, the four learning types, the four SageMaker inference modes, and the ML lifecycle stages.
| Type | What it learns from | AWS examples |
|---|---|---|
| Supervised | Labeled data → classification or regression | Comprehend (sentiment), Rekognition, Fraud Detector |
| Unsupervised | Unlabeled data → clustering, anomaly detection | SageMaker K-Means, Random Cut Forest |
| Reinforcement | Reward signals from an environment | DeepRacer, SageMaker RL |
| Self-supervised | Generates its own labels from unlabeled data | How foundation models are pre-trained |
| Type | Latency | Payload | Use case |
|---|---|---|---|
| Real-time | Low (ms) | Small | Persistent endpoint, always-on traffic (chatbots, fraud) |
| Serverless | Low–Med | Small | Intermittent / unpredictable traffic, scales to zero |
| Asynchronous | High | Up to 1 GB | Long processing (up to 1 hr), large payloads, queued |
| Batch Transform | Highest | Very large | Bulk offline jobs, no endpoint, scheduled |
Business problem framing → Data collection (S3, Kinesis) → Data prep (Data Wrangler, Glue) → Feature engineering (Feature Store) → Training (SageMaker Training Jobs) → Evaluation → Deployment (SageMaker Endpoints) → Monitoring (Model Monitor, Clarify).
SageMaker Clarify). Variance = sensitivity to training-data fluctuations.Automatic Model Tuning. Parameters are the weights/biases learned during training.Domain 2 is 24% of the exam. The big building blocks: foundation models, tokens, embeddings, vector DBs, and the flavours of generative architectures.
A foundation model is a large model pre-trained on massive datasets via self-supervised learning, adaptable to many downstream tasks. Fine-tuning a FM is far cheaper than training from scratch. On Bedrock you'll find:
| Concept | Definition |
|---|---|
| Tokens | Sub-word chunks (~4 chars in English). Bedrock pricing is per input + output token. |
| Embeddings | Numerical vector representations of text/images. Used for semantic search & RAG. Services: Titan Embeddings, Cohere Embed. |
| Vector database | Stores embeddings for similarity search. |
| Context window | Max tokens a model processes at once. |
| Temperature | Controls randomness — 0 = deterministic, 1 = creative. |
pgvector extension.DynamoDB is not a vector database. Don't pick it for embedding storage.| Architecture | What it's used for |
|---|---|
| Transformers | Attention-based — backbone of most modern LLMs. |
| GANs | Generative Adversarial Networks — image generation, two-network duel. |
| VAEs | Variational Autoencoders — data generation from a learned latent space. |
| Diffusion models | Image generation (Stable Diffusion) — denoise from random noise. |
Hallucinations (confident but false — mitigate with RAG/grounding/Guardrails), knowledge cutoff (mitigate with RAG), cost (long contexts add up), latency (bigger models = slower), bias (reflects training data), non-determinism (same input → different output when temperature > 0).
Bedrock is fully managed and serverless, providing single-API access to multiple FMs with no infrastructure to manage. If a question says "no infrastructure", "fully managed", or "API-only", the answer is almost always Bedrock.
| Aspect | Amazon Bedrock | SageMaker JumpStart |
|---|---|---|
| Infrastructure | Fully managed, serverless | You deploy on YOUR SageMaker |
| Access | Single API call | Deploy endpoint first |
| Control | Limited, abstracted | Full (instances, hyperparams) |
| Use case | Fast, hands-off GenAI | Deep customization needed |
Customer prompts and outputs on Bedrock are NOT used to train base FMs. Your data stays yours. Bedrock is HIPAA-eligible, SOC compliant, GDPR-ready, PCI DSS supported.
SageMaker is the all-in-one ML platform — every lifecycle stage has a SageMaker tool. The exam tests whether you can match a stage to the right service.
| Stage | Service | What it does |
|---|---|---|
| Data prep | Data Wrangler | Visual data prep — 300+ built-in transforms, no code needed. |
| Feature engineering | Feature Store | Centralized store for features — share across training & inference. |
| Training | Training Jobs | Managed training on the instance type you pick. |
| Hyperparameter tuning | Automatic Model Tuning | Bayesian / random search across hyperparameter ranges. |
| Pre-built models | JumpStart | Deploy open-source FMs & hundreds of pretrained models. |
| Deployment | Endpoints | Real-time / serverless / async / batch — see Domain 1 panel. |
| Bias & explainability | Clarify | Detects bias in data & trained models. Provides SHAP explainability. |
| Drift monitoring | Model Monitor | Detects data drift & concept drift in production. |
| Human-in-loop | Augmented AI (A2I) | Sends low-confidence predictions to human reviewers. |
| Model documentation | Model Cards | You document YOUR custom models — training data, intended use, etc. |
Domain 3 is the highest-weighted domain at 28%. Master the customization spectrum and prompt engineering — they together cover most of the questions.
| Method | Cost | Use when… |
|---|---|---|
| Prompt engineering | Free | Simple tasks, no special data |
| RAG | Cheap | Private/recent factual data, reduce hallucinations |
| Fine-tuning (instruction) | Moderate | Specific style/format/tone, with labeled examples |
| Domain adaptation FT | Moderate | Limited domain-specific labeled data |
| Continued pre-training | Expensive | Lots of UNLABELED domain data |
| Train from scratch | V. expensive | Rarely needed — only for fundamentally new architectures |
Bedrock Knowledge Bases handles all 5 steps automatically. Connects to S3, Confluence, Salesforce, SharePoint, web crawler.
Prompt injection (malicious user input overrides system instructions), prompt leaking (tricking the model into revealing its system prompt), jailbreaking (bypassing guardrails), poisoning (harmful data during fine-tuning), hijacking (taking control of model behaviour).
Mitigation: Bedrock Guardrails, input validation, content filtering.
| Parameter | Effect |
|---|---|
| Temperature (0–1) | Randomness. Low = factual, deterministic. High = creative. |
| Top-p (nucleus) | Considers tokens whose probabilities sum to p. |
| Top-k | Considers the top-k most likely tokens. |
| Max tokens | Caps the output length. |
| Stop sequences | Strings that halt generation when emitted. |
Agents enable LLMs to take actions, not just generate text. Components:
The exam tests whether you can match an AI use case to the right pre-built AWS service. Memorize the Q variants and the classic managed-AI line-up.
| Use case | Service |
|---|---|
| Sentiment, entity, key-phrase, language detection in text | Amazon Comprehend |
| Image & video analysis (faces, labels, text in images, moderation) | Amazon Rekognition |
| Speech → text | Amazon Transcribe |
| Text → speech | Amazon Polly |
| Translation between languages | Amazon Translate |
| Extract text + tables + forms from documents | Amazon Textract |
| Detect online fraud (account takeover, payment fraud) | Amazon Fraud Detector |
| Personalised recommendations | Amazon Personalize |
| Forecast time-series data | Amazon Forecast (now via SageMaker Canvas) |
| Conversational chatbots | Amazon Lex |
| Healthcare entity extraction (HIPAA-eligible) | Amazon Comprehend Medical |
| Industrial vision quality inspection | Amazon Lookout for Vision |
Domain 4 is 14%. Mostly conceptual — know the 8 responsible-AI principles, the transparency vs explainability distinction, and which AWS service implements each.
| Aspect | Transparency | Explainability |
|---|---|---|
| Focus | The whole system | A specific prediction |
| Question answered | What data, what model, what limits, who built it? | Which features contributed to this output and how? |
| Form | Documentation (Model Cards, Service Cards) | Feature attribution (SHAP, LIME) |
| Memory aid | Visible from outside (glass box) | Model explains its reasoning |
Data drift: input distribution changes (new product categories appear). Concept drift: the relationship between inputs and outputs changes (what counts as fraud evolves). Mitigation: SageMaker Model Monitor → detect drift → trigger retraining.
Sources: data, algorithmic, sampling, confirmation, measurement bias.
Mitigation: diverse balanced datasets, audit across demographic slices, SageMaker Clarify, fairness metrics (demographic parity, equalized odds), continuous post-deployment monitoring.
Domain 5 is 14%. AWS reuses its standard security stack (IAM, KMS, VPC, CloudTrail) and adds a few AI-specific pieces (Bedrock Model Invocation Logging, Guardrails, the GenAI Scoping Matrix).
| Service | What it does for AI |
|---|---|
| IAM | Least privilege; use roles (not long-term keys) for Bedrock/SageMaker/Q. SCPs for org guardrails. IAM Identity Center → user identity for Q Business. |
| KMS | Encryption at rest for S3, EBS, SageMaker volumes. Customer-managed keys (CMKs) supported. |
| Secrets Manager | Stores API keys, DB credentials. |
| VPC + VPC Endpoints (PrivateLink) | Access Bedrock / SageMaker / Q WITHOUT going over the public internet. |
| SageMaker VPC mode | Run training/inference inside your own VPC. |
| CloudTrail | Logs all API calls — who invoked which model, when. |
| CloudWatch | Metrics & logs (latency, errors, token usage). |
| AWS Config | Tracks configuration changes & compliance. |
| Bedrock Model Invocation Logging | Logs PROMPTS and RESPONSES to S3 / CloudWatch — required if the question says "audit prompts AND responses". |
| Macie | Detects PII / sensitive data in S3 training corpora. |
| Lake Formation / Glue Catalog / DataZone | Fine-grained access & governance over data lakes. |
| Scope | What it is | Control vs responsibility |
|---|---|---|
| Scope 1 | Consumer apps (e.g., public ChatGPT) | Least control, least responsibility |
| Scope 2 | Enterprise SaaS with GenAI features | ↓ |
| Scope 3 | Pre-trained FMs (Bedrock on-demand) | ↓ |
| Scope 4 | Fine-tuned FMs with your data | ↓ |
| Scope 5 | Self-trained models from scratch | Most control, most responsibility |
Prompt injection (mitigate with Bedrock Guardrails), data leakage through prompts (users pasting confidential info), model theft / extraction (adversaries copying via API), training-data poisoning (corrupted data biases the model), membership inference (determining if specific data was in training set), PII exposure in outputs.
AWS Artifact (SOC, ISO, HIPAA, PCI DSS, GDPR reports), Audit Manager, AWS Config, SCPs (org-wide), AWS Budgets (cost alerts), Cost Explorer, tagging by project, provisioned vs on-demand, Bedrock Batch (50% cheaper).
Easy points if you memorize them. The exam loves single-word triggers ("translation" → BLEU; "summarization" → ROUGE).
| Metric | Used for |
|---|---|
| BLEU | Translation quality |
| ROUGE | Summarization quality |
| BERTScore | Semantic similarity using BERT embeddings |
| Perplexity | How well a language model predicts a sample (lower = better) |
| Human evaluation | Gold standard, expensive — best for nuanced quality |
| Bedrock Model Evaluation | Automated + human evaluation service for comparing FMs |
| Metric | What it measures |
|---|---|
| Accuracy | Overall correctness — misleading on imbalanced data. |
| Precision | Of predicted positives, how many were correct? Use when false positives are costly. |
| Recall | Of actual positives, how many did we catch? Use when false negatives are costly (e.g., disease detection). |
| F1 | Harmonic mean of precision & recall. |
| AUC-ROC | Trade-off across thresholds. |
Capability (reasoning / coding / multilingual / vision), cost per token, latency, context window, modalities (text vs multimodal), customization support, regional availability.
Beyond model quality: task completion rate, user satisfaction (CSAT), cost per inference, time to resolution, conversion rate, error/hallucination rate, ROI.
Click any component below to see what it does, where it sits in a typical Bedrock + RAG production stack, and which AIF-C01 questions it answers. Top-down: data sources → embedding pipeline → vector store → FM serving → app, with security and observability cutting across every layer.
30 questions in AIF-C01 style, mixed across all five domains. Click to expand. Filter by difficulty.
Bedrock Knowledge Bases handles all 5 steps automatically.DynamoDB is NOT a vector database — don't pick it for embedding storage.Retrieve and Generate from a Bedrock model. RAG (not fine-tuning) because the data is private and changes frequently. Add Bedrock Guardrails for safety and use VPC endpoints to keep traffic private.15 multiple-choice questions in AIF-C01 style. Timed at 22 minutes (same per-question pace as the real 65-question, 90-minute exam). Pass mark: 70%. You'll see an explanation after each question.
A mix of easy, medium, and hard questions sampled across all five AIF-C01 domains: AI/ML Fundamentals, Generative AI, Applications of FMs, Responsible AI, and Security. You'll get an explanation after each question and a full breakdown at the end.
—
Six guided labs you can run in the AWS Free Tier (most use Bedrock — request model access first in the console). Console click-paths and AWS CLI snippets side by side. The console click-paths are stable; the CLI snippets show the equivalent API calls.
Enable model access for Anthropic Claude on Bedrock and send your first prompt from the Playground and the CLI.
Bedrock → Model access → Manage model access → request Anthropic Claude (approval is usually instant).Bedrock → Playgrounds → Chat → pick Claude → "Explain RAG in 3 sentences".aws bedrock-runtime invoke-model --model-id anthropic.claude-3-haiku-20240307-v1:0 --body '{"messages":[{"role":"user","content":"Explain RAG"}],"max_tokens":256,"anthropic_version":"bedrock-2023-05-31"}' out.jsonout.json — note the input_tokens and output_tokens fields (this is what you're billed on).Index a folder of PDFs in S3, point a Bedrock Knowledge Base at the bucket, and query it.
s3://your-aiprep-bucket/docs/.Bedrock → Knowledge Bases → Create. Pick S3 as data source, OpenSearch Serverless as vector store, Titan Embeddings as the embedding model.Test knowledge base → Retrieve and generate and ask a question grounded in the PDFs.Create an agent that calls a Lambda function as an action group — multi-step task with real action.
get_weather(city) that returns a stub response.Bedrock → Agents → Create. Pick Claude as the FM, paste the OpenAPI schema, link the Lambda.Block sensitive topics and redact PII from FM outputs.
Bedrock → Guardrails → Create guardrail.InvokeModel calls to apply it at runtime.Run a Clarify processing job on a tabular dataset to surface bias and feature importance.
s3://your-aiprep-bucket/clarify/.facet_name=sex and label_values_or_threshold=1.Deploy a model, capture inference data, and create a monitoring schedule.
data_capture_config enabled (writes to S3).DefaultModelMonitor.Eight production patterns AWS likes to test as scenarios. Read each as: what the customer wants → which AWS services solve it → why.
Grab the PDFs to revise offline, then memorize the decision patterns below — they'll solve a big chunk of the exam by pattern-matching alone.
| If the question mentions… | Answer is likely… |
|---|---|
| "No infrastructure" / "fully managed" / "API" | Amazon Bedrock |
| "Full control" / "your own infrastructure" | SageMaker JumpStart |
| "Private data" / "company documents" / "no retraining" | RAG / Bedrock Knowledge Bases |
| "Specific style/tone" + "labeled examples" | Instruction fine-tuning |
| "Lots of UNLABELED domain data" | Continued pre-training |
| "Multi-step task" + "take actions" + "API calls" | Bedrock Agents |
| "Detect bias" / "feature importance" / "SHAP" | SageMaker Clarify |
| "Low confidence" + "human review" | Amazon A2I |
| "Data drift" / "model degradation" | SageMaker Model Monitor |
| "Detect PII in S3" | Amazon Macie |
| "Block topics, filter PII in outputs" | Bedrock Guardrails |
| "Audit prompts AND responses" | Bedrock Model Invocation Logging |
| "Without public internet" | VPC Endpoints / PrivateLink |
| "Vector storage / semantic search" | OpenSearch Serverless (NOT DynamoDB) |
| "Translation evaluation" | BLEU |
| "Summarization evaluation" | ROUGE |
| "Code assistant in IDE" | Amazon Q Developer |
| "Contact center agent assist" | Amazon Q in Connect |
| "Enterprise data assistant" | Amazon Q Business |
| "Detect online fraud" | Amazon Fraud Detector |
| "OCR / extract tables & forms from PDFs" | Amazon Textract |
| "Image labels / face detection / moderation" | Amazon Rekognition |
| "Speech to text" | Amazon Transcribe |
| "Text to speech" | Amazon Polly |
| "Translate text" | Amazon Translate |
| "Personalization / recommendations" | Amazon Personalize |
| "Conversational chatbot" | Amazon Lex |
| Mode | Latency | Payload | Pick when… |
|---|---|---|---|
| Real-time | Low (ms) | Small | Always-on, persistent endpoint, low-latency required |
| Serverless | Low–Med | Small | Intermittent / unpredictable traffic, scales to zero |
| Asynchronous | High | Up to 1 GB | Long processing (up to 1 hr), large payloads, queued |
| Batch Transform | Highest | Very large | Bulk offline jobs, no endpoint, scheduled |
| Method | Cost | Use when |
|---|---|---|
| Prompt engineering | Free | Simple tasks, no special data |
| RAG | Cheap | Private/recent data, reduce hallucinations |
| Fine-tuning (instruction) | Moderate | Specific style/format/tone, labeled examples |
| Domain adaptation FT | Moderate | Limited domain-specific labeled data |
| Continued pre-training | Expensive | Lots of UNLABELED domain data |
| Train from scratch | V. expensive | Rarely needed |