STEP 1 · Study

AIP-C01 StudyGenerative AI Developer – Professional

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

Foundation Model Integration, Data Management, and Compliance

31%

Learning goals

Choose FMs, integration patterns and deployment methods from business requirements and constraints, build PoCs on Bedrock, and align designs with the Well-Architected Generative AI Lens
Select FMs through benchmarks and evaluation, and design configurations that switch models or providers without code changes (AppConfig, Lambda, API Gateway) plus resilience with cross-Region inference and circuit breakers
Build validation and processing for data passed to FMs (Glue Data Quality, multimodal, conversation formats), and design vector stores (OpenSearch, Aurora pgvector, S3 Vectors, Bedrock knowledge bases) with metadata and update mechanisms
Design retrieval with chunking, embeddings, hybrid search, reranking and query expansion, and govern prompts with Prompt Management, templates, guardrails and prompt flows

Key points

Design approach: decide requirements first (quality, latency, cost, data location, regulation), evaluate multiple Bedrock models in a small PoC, then move to production design. Standardize shared components (prompt templates, guardrails, logging).
Switching models: do not hard-code model IDs or parameters; manage them in configuration such as AWS AppConfig. The Converse API lets you call different models with a common format.
Resilience: prepare for Regional capacity shortages and throttling with cross-Region inference, and design graceful degradation such as Step Functions retries and circuit breakers and falling back to a smaller model.
Customization: start with prompts and RAG. If that is not enough, use fine-tuning (parameter-efficient methods such as LoRA) or distillation. Version custom models so you can roll back on failure.
Data processing: validate quality before passing data to FMs (Glue Data Quality, Lambda), and process images, audio and tables (Transcribe, Bedrock Data Automation, multimodal models). Pass conversations in role format (user, assistant).
Vector stores: OpenSearch for hybrid search or large scale, pgvector on Aurora or RDS for existing PostgreSQL, S3 Vectors for storing large volumes cheaply. Narrow retrieval with metadata (date, author, classification). Reflect document changes with incremental sync.
Retrieval design: chunking (fixed, hierarchical, semantic, custom Lambda), embedding model and dimensions, hybrid search, reranking, query rewriting and decomposition.
Prompts: set role and constraints in the system prompt, manage templates and versions with Bedrock Prompt Management, and apply approval and auditing. Split complex tasks into steps and wire order and branches with Prompt Flows.

Workplace uses

SalesFor a proposal assistant, the team wants to compare and switch models monthly. Model IDs are managed in AppConfig and called via the Converse API, so new models can be adopted without code changes.
General AffairsIn an internal policy RAG, old pre-revision policies appear in answers. Add effective-date metadata to documents to filter retrieval, and run knowledge base sync automatically when S3 changes.
AccountingBefore an FM reads invoice images and tables, Lambda validates amount and date formats, and invoices with missing fields are routed to staff for review.

Exam objectives (from the exam guide)

1.1 要件を分析し、GenAI ソリューションを設計する。
1.2 FM を選定して設定する。
1.3 FM 消費のためのデータ検証と処理パイプラインを実装する。
1.4 ベクトルストアソリューションを設計して実装する。
1.5 FM 拡張のための検索メカニズムを設計する。
1.6 FM インタラクションのためのプロンプトエンジニアリング戦略とガバナンスを実施する。

Learning resources

Learn in detail (6 topics with mini quizzes) →Practice this domain (93 questions) →Next domain ↓

Implementation and Integration

26%

Learning goals

Build agents and multi-agent systems with memory, tools and human approval using Strands Agents, Bedrock Agents, AgentCore (Runtime, Memory, Gateway, Identity) and MCP
Choose FM deployment methods by requirements among on-demand, Provisioned Throughput, SageMaker AI endpoints and Custom Model Import
Connect securely to existing business systems with API Gateway, EventBridge, Lambda and PrivateLink, and design a generative AI gateway and data residency requirements
Implement synchronous, asynchronous and streaming API calls, retries with backoff and model routing, and speed up development with Amplify, Q Developer and similar

Key points

Agents: the model plans, calls tools, looks at results and decides the next action (ReAct). Strands Agents is an open-source SDK for building in code, Bedrock Agents is managed, and AgentCore is the production runtime foundation (Runtime, Memory, Gateway, Identity, Observability).
Tool integration: expose tools in a standard way with MCP. Define tool inputs with JSON Schema and include timeouts, retries and error handling. Add human approval (such as return of control) for important actions.
Multi-agent: a supervisor agent delegates work to specialist agents. Protocols such as A2A exist for agent-to-agent communication. Prevent runaway behavior with iteration limits and timeouts.
Deployment: on-demand for pay-per-use, Provisioned Throughput for fixed throughput, SageMaker AI to manage instances and containers, Custom Model Import for open models tuned elsewhere.
Enterprise integration: build a generative AI gateway on API Gateway that centrally manages authentication, throttling and usage plans. Integrate asynchronously with EventBridge or SQS, and keep traffic private with PrivateLink.
API calls: InvokeModel and Converse, streaming (ConverseStream) to show responses immediately, batch inference for high volume. Retry throttling with exponential backoff and jitter.
Routing: use small and large models depending on question difficulty or type (Intelligent Prompt Routing or a custom router).
Developer productivity: Amplify for the frontend, Q Developer for code generation and explanation, Bedrock Prompt Flows to compose flows. Troubleshoot with CloudWatch Logs and X-Ray.

Workplace uses

SalesA deal-preparation agent uses the CRM API and the internal price list as MCP tools to draft proposals. Discount approvals are confirmed by a person before execution.
General AffairsEmbed an equipment-request chat in the internal portal. Set per-department usage limits in API Gateway and stream responses so waiting feels shorter.
AccountingGenerate descriptions for 50,000 month-end journal entries with batch inference instead of synchronous on-demand calls, keeping costs down and finishing by the next morning.

Exam objectives (from the exam guide)

2.1 エージェンティック AI ソリューションとツール統合を実装する。
2.2 モデルデプロイ戦略を実施する。
2.3 エンタープライズ統合アーキテクチャを設計し実装する。
2.4 FM API 統合を実装する。
2.5 アプリケーション統合パターンと開発ツールを実装する。

Learning resources

Learn in detail (5 topics with mini quizzes) →Practice this domain (78 questions) →Next domain ↓

AI Safety, Security, and Governance

20%

Learning goals

Build Bedrock Guardrails (content filters, denied topics, word filters, sensitive information, grounding, Automated Reasoning checks) and prompt-attack defenses in layers on both input and output
Protect FM environments and data and preserve privacy with VPC endpoints, IAM, KMS, Macie and Comprehend PII detection
Meet regulatory and governance requirements with CloudTrail, Config, invocation logging, model cards and lineage tracking
Build transparency (showing sources), fairness evaluation and responsible AI policies into application design and operations

Key points

Input safety: detect jailbreaks and prompt injection with the Guardrails prompt-attack filter. Treat external documents and tool results as untrusted input and keep them separate from instructions.
Output safety: content filters, denied topics and sensitive information filters (mask or block). Reduce hallucinations with contextual grounding checks and rule-based Automated Reasoning checks.
Defense in depth: stack input validation → guardrails → least-privilege tools → output validation → human review. Guardrails can also be used separately from the model via the ApplyGuardrail API.
Data protection: keep traffic private with VPC endpoints (PrivateLink) and encrypt with KMS customer managed keys. Bedrock does not use inputs and outputs to train models and does not share them with model providers.
Privacy: find sensitive data in S3 with Macie, and detect and mask PII with Comprehend or Guardrails. Keep personal information out of logs too.
Governance: record API actions with CloudTrail, inputs and outputs with model invocation logging, and configuration rules with Config. Document purpose and limitations in model cards, and keep data provenance (citations) traceable.
Responsible AI: show the basis (citations) for answers and disclose that content is AI-generated. Measure quality differences across groups through evaluation, and prevent off-policy use with guardrails and approvals.

Workplace uses

SalesIn a customer-facing chat, set denied topics and content filters so it does not disparage competitors or promise prices, and attach citations of source materials to answers.
General AffairsIn an employee consultation assistant, mask names and phone numbers in consultations with the sensitive information filter so they do not remain in invocation logs either.
AccountingIn an assistant that answers tax questions, verify answers with Automated Reasoning checks where rule-based judgment is needed, and stop content not in retrieved documents with grounding checks.

Exam objectives (from the exam guide)

3.1 入力と出力の安全コントロールを実装する。
3.2 データセキュリティとプライバシーコントロールを実装する。
3.3 AI ガバナンスとコンプライアンスのメカニズムを実装する。
3.4 責任ある AI の原則を実装する。

Learning resources

Learn in detail (4 topics with mini quizzes) →Practice this domain (60 questions) →Next domain ↓

Operational Efficiency and Optimization for GenAI Applications

12%

Learning goals

Lower costs while keeping quality through token reduction, model selection, prompt caching, response caching and batch inference
Improve latency and throughput with streaming, retrieval optimization, securing throughput and parameter tuning
Monitor FM applications, tools and vector stores with CloudWatch generative AI observability, invocation logging, AgentCore Observability and X-Ray
Detect and handle generative-AI-specific failures such as token limits, context truncation and throttling

Key points

Token efficiency: shorten system prompts, limit the number and length of RAG chunks, and summarize conversation history. Set the maximum output tokens.
Model selection: choose the smallest, cheapest model that meets quality in evaluation. Route easy questions to small models and hard ones to large models.
Caching: use prompt caching for repeated prompt prefixes. For answers to the same question, use semantic caching (embedding similarity) in ElastiCache or similar to reduce FM calls.
Latency: make time-to-first-token feel shorter with streaming, and use latency-optimized inference or smaller models. Parallelize what can be parallelized.
Throughput: for throttling, use cross-Region inference, Provisioned Throughput, quota increases and backoff. Use batch inference for non-urgent work.
Monitoring: track token volume, latency, errors and throttling in CloudWatch, inputs and outputs with invocation logging, and agent steps and tools with AgentCore Observability. Also monitor vector store latency and capacity.
Generative-AI-specific failures: truncation beyond context length, stopping mid-output at max_tokens, tool-failure loops, retrieval drift, and quality changes from model updates.

Workplace uses

SalesMany similar product questions come in, so add a semantic cache that returns past answers when question embeddings are close, reducing FM calls and wait time.
General AffairsIn response to complaints that the internal assistant is slow, stream responses and reduce retrieved chunks from 10 to 4, confirming quality through evaluation.
AccountingThrottling increases at month-end, so use a cross-Region inference profile and route non-urgent work to batch inference. Notify with CloudWatch alarms.

Exam objectives (from the exam guide)

4.1 コスト最適化とリソース効率化の戦略を実施する。
4.2 アプリケーションのパフォーマンスを最適化する。
4.3 GenAI アプリケーションのモニタリングシステムを実装する。

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (36 questions) →Next domain ↓

Testing, Validation, and Troubleshooting

11%

Learning goals

Evaluate FM outputs from multiple angles by combining automatic metrics, LLM-as-a-judge, human evaluation and user feedback
Measure RAG retrieval quality, answer faithfulness and agent task success rates, and validate before and after deployment
Isolate and fix context-length, API integration, prompt and retrieval problems from logs and traces

Key points

Evaluation dimensions: correctness, relevance, faithfulness, completeness, harmfulness and tone. With reference answers, use ROUGE, BERTScore and similar; without them, LLM-as-a-judge or human evaluation.
Bedrock evaluations: model evaluation (automatic, human, LLM-as-a-judge) and knowledge base evaluation (retrieval relevance, answer correctness).
User perspective: collect feedback such as thumbs up/down, and add low-rated conversations to the evaluation dataset.
Agent evaluation: check task completion rate, correctness of tool choice, step count and cost, using traces.
Deployment validation: regression tests on the evaluation dataset, shadow or canary comparisons, and rollback when quality drops.
Troubleshooting: input too long (context limit) → summarize or chunk. ValidationException → request format. ThrottlingException → backoff and quotas. Unstable answers → temperature and prompt. Retrieval misses → chunking, embeddings and hybrid search.

Workplace uses

SalesBefore switching to a new model, compare it with the current model using LLM-as-a-judge on an evaluation dataset of 200 typical sales questions, confirming correctness does not drop.
General AffairsInvestigating questions with many “answer not found” low ratings shows policy tables were split by chunking. Revise chunking for content with tables and confirm the improvement with knowledge base evaluation.
AccountingSummaries of long contracts get cut off. The stop reason was max_tokens, so revise the output limit, and for documents with overly long input, summarize each chapter first and then combine.

Exam objectives (from the exam guide)

5.1 GenAI の評価システムを実装する。
5.2 GenAI アプリケーションをトラブルシューティングする。

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (33 questions) →Try the mock exam →

When you finish reading

Practice all domains (300 questions)Mock exam (75 questions · 180 min)AIP-C01 home

Notes (unofficial, original questions, no pass guarantee)

The questions on this site are original practice questions written by this site based on publicly available skills outlines and exam guides. They are not actual exam questions.
Amazon Web Services, AWS and related marks are trademarks of Amazon.com, Inc. or its affiliates. This site is not affiliated with or endorsed by AWS.
This content is a study aid and does not guarantee a pass. Exam content, format and prerequisites may change without notice; always check the latest information on the official page before taking the exam.
Your accuracy on the mock exam cannot be converted to the official (scaled) score and does not predict a pass or fail.
Progress is saved only in this browser (localStorage) unless you sign in with Google to sync it. You can delete it from the Cookie & storage settings page (Japanese).