AWS GenAI Developer Pro study plan

Map each of the six exam domains to its weight and write a one-paragraph summary of what Domain 1 covers before touching any code.

Build a working Amazon Bedrock Converse API chat loop in Python that streams tokens to the terminal and handles a mid-stream error without crashing.

Compare model families available in Bedrock on context window, modality support, and streaming behavior, and record the comparison in a table.

Configure inference parameters (temperature, top-p, max tokens, stop sequences) and document how each changes output on the same prompt.

Implement tool use (function calling) with at least two tools and validate that the model returns structured arguments you can execute.

Add retry with exponential backoff and jitter around throttling and transient errors, and prove it works with a simulated failure.

Design a prompt template layer that separates system instructions, retrieved context, and user input into distinct message roles.

Instrument the client to log latency and token counts per request.

Build a retrieval pipeline that chunks documents, generates embeddings, and stores vectors in a vector store.

Tune chunk size and overlap and measure retrieval quality at two settings using a small hand-labeled question set.

Implement a RAG prompt that cites source chunks and instructs the model to refuse when context is insufficient.

Add a reranking or filtering step and quantify its effect on answer relevance.

Connect the RAG app to Amazon Bedrock Knowledge Bases and compare managed retrieval against your custom pipeline.

Define a continuous evaluation harness that scores faithfulness, relevance, and answer correctness on a fixed test set.

Explain when fine-tuning beats RAG and write a decision rule you can defend in a design review.

Prepare a small instruction dataset in the format required for fine-tuning.

Run a fine-tuning or model customization job and capture training and validation metrics.

Compare base versus tuned model outputs on held-out prompts and quantify the improvement.

Set up automated evaluation that runs on every model or prompt change and blocks regressions.

Document a rollback plan for a model version that fails evaluation.

Configure IAM least-privilege policies so the application role can invoke only the specific models and knowledge bases it needs.

Encrypt data at rest and in transit and verify that no prompt or response is logged in plaintext where it should not be.

Implement guardrails against prompt injection and verify they block a set of adversarial test inputs.

Apply PII detection and redaction to both inbound prompts and outbound responses.

Instrument the application with metrics, traces, and dashboards covering latency, error rate, token usage, and cost per request.

Configure content filters and denied-topic guardrails and test them against a policy test suite.

Model cost for a projected workload and identify two concrete levers to reduce it.

Write an incident runbook for a model outage or guardrail false-positive surge.

Take two full-length timed practice exams and score above 80 percent on each.

Review every missed question and map it back to the owning domain.

Re-read notes for the two weakest domains and rebuild any lab you could not explain from memory.

Restate each domain's core services and trade-offs aloud without notes.