AWS Certified Generative AI Developer - Professional (AIP-C01) - Practice Plan¶
An 8-week plan assuming 6-8 hours a week. Compress to 5 weeks if you already build production GenAI applications on Bedrock daily; stretch to 12 if Bedrock is new to you.
This is a Professional-tier exam. It assumes you can already write application code, call APIs, and reason about IAM. If those are shaky, spend two weeks on AWS Certified AI Practitioner first, which is background rather than a prerequisite.
Prerequisites worth having
- Comfortable with Python and the AWS SDK
- Working IAM knowledge: identity vs resource policies, least privilege
- Some exposure to vector search concepts (see embeddings and vector search)
Week 1 - Foundation models and the Bedrock surface¶
Goal: know what the platform offers before deciding how to use it.
- Read fact-sheet.md end to end. Note the domain weights.
- Read notes/01-foundation-models-data-compliance.md
- Read notes/bedrock-platform-deep-dive.md, the single most-tested service surface
- Hands-on: invoke three different foundation models on the same prompt via the Bedrock API. Compare output, latency, and cost per 1K tokens
- Write down, in your own words, when you would pick each model family
Checkpoint: you can explain model selection in terms of capability, latency, and cost without looking anything up.
Week 2 - Prompt engineering and management¶
- Read notes/prompt-engineering-and-management.md
- Read the concept page on prompt engineering
- Hands-on: take a task that fails with a naive prompt and fix it with structure, few-shot examples, and explicit output format
- Hands-on: version two prompts and compare them on the same 10 inputs
- Note where prompt engineering stops being enough. That boundary is tested
Checkpoint: you can articulate the prompt engineering / RAG / fine-tuning decision and name the constraint that selects each.
Week 3 - RAG architecture, part 1¶
The largest single topic on the exam.
- Read notes/rag-architecture-deep-dive.md
- Read rag-explained and fine-tuning vs RAG
- Hands-on: build a Knowledge Base over a document set you know well
- Deliberately break it: oversized chunks, no overlap, no metadata. Observe how the failure looks from the answer side
Checkpoint: given a bad answer, you can tell whether the fault is retrieval or generation.
Week 4 - RAG architecture, part 2, and data management¶
- Compare vector store options and their trade-offs. See vector database comparison and the decision matrix
- Hands-on: add metadata filtering and measure the precision improvement on 10 questions
- Study data handling: where prompts and completions are stored and logged, and what that means for regulated data
- Re-read the compliance half of notes/01-foundation-models-data-compliance.md
Checkpoint: you can design a RAG pipeline for a regulated workload and justify each component.
Week 5 - Implementation, integration, and agents¶
- Read notes/02-implementation-integration.md
- Read notes/agentic-ai-systems.md
- Read the concept pages on tool use and agentic loops
- Hands-on: build an agent with two action groups whose purposes are genuinely close. Watch it misroute, then fix it purely by rewriting descriptions
- Hands-on: add a third action group that calls an external API
Checkpoint: you can debug tool selection without reaching for a bigger model.
Week 6 - Safety, security, and governance¶
- Read notes/03-ai-safety-security-governance.md
- Read guardrails and safety
- Hands-on: configure a Guardrail with PII filters on both input and output, and confirm both directions actually block
- Hands-on: write an IAM policy that permits one model and denies another for a specific role. Test it
- Practice stating the difference between a Guardrail control and an IAM control in one sentence
Checkpoint: you never confuse content filtering with access control. This is worth several marks.
Week 7 - Optimization, testing, and validation¶
- Read notes/04-operational-efficiency-optimization.md
- Read notes/05-testing-validation-troubleshooting.md
- Read evals for LLMs and prompt caching
- Hands-on: build a small evaluation set (20 questions with known-good sources) and score two configurations against it
- Work through the cost model: On-Demand vs Provisioned Throughput vs batch, for three different duty cycles
- Hands-on: measure the latency difference streaming makes to time-to-first-token
Checkpoint: you can decide On-Demand / Provisioned / batch from a described duty cycle, and you can describe a defensible evaluation.
Week 8 - Consolidation and exam readiness¶
- Work all of scenarios.md. For each, write why every distractor fails
- Read strategy.md and internalize the trap list
- Skim notes/aws-services-mapping.md as a reverse index: capability to service
- Re-read notes/bedrock-platform-deep-dive.md and notes/rag-architecture-deep-dive.md
- Take a timed practice run. Hold to 2.4 minutes per question
- Review every miss and classify it: knowledge gap, misread, or timing
Do not start new material in the final three days. Consolidate.
Ongoing habits¶
- Keep a running list of every question you get wrong and the reason. Patterns show up fast.
- Any time you read "least operational overhead" or "most cost-effective," stop and note which one it is. The exam turns on that distinction constantly.
- Build things. This exam rewards people who have actually shipped a RAG pipeline and debugged an agent over people who have only read about them.