AWS Generative AI Developer – Professional

Start with the largest domain. Learn to turn business constraints into an architecture and validate model, data, retrieval, and prompt choices with representative evidence.

• Map use cases to quality, latency, context, safety, Region, and cost requirements
• Compare supported Bedrock models with a representative evaluation set
• Know when prompting, RAG, customization, or a smaller specialized model fits
• Build modality-aware validation and processing pipelines
• Understand embeddings, dimensions, vector stores, metadata, and indexing
• Tune parsing, fixed-size or structure-aware chunking, overlap, and synchronization
• Compare semantic, hybrid, filtered, and reranked retrieval
• Use Bedrock Knowledge Bases for retrieval, generation, and citations
• Use Prompt Management variables, variants, testing, and versions
• Design model abstraction, graceful degradation, and compliant regional routing

Implement production interfaces rather than isolated playground prompts. Separate probabilistic planning from deterministic authorization, workflow control, and tool execution.

• Use AWS SDKs and the appropriate Bedrock inference interface
• Design streaming responses for interactive experiences
• Use SQS or EventBridge for asynchronous, decoupled workloads
• Apply bounded retries, backoff with jitter, rate limits, and fallbacks
• Build stable model gateways and configuration-driven routing
• Define agent tools with strict schemas and structured errors
• Validate resource scope, authorization, range, and idempotency in tool code
• Use Step Functions for stopping conditions, timeouts, approval, and rollback
• Integrate human expertise into high-impact workflows
• Version and test prompts, models, retrieval, policies, APIs, and infrastructure in CI/CD

Treat model input, retrieved content, tool arguments, and output as separate trust boundaries. Implement controls in code and cloud policy instead of depending on natural-language instructions alone.

• Configure Guardrails for relevant content, topic, word, sensitive-data, and grounding policies
• Test prompt injection, jailbreak, indirect injection, and unsafe-output paths
• Apply least-privilege IAM to models, prompts, guardrails, knowledge bases, tools, and data
• Use KMS-backed encryption and Secrets Manager where appropriate
• Use Bedrock VPC endpoints with PrivateLink and restrictive endpoint policies
• Protect tenant and business-unit authorization in trusted application logic
• Track data and model sources, versions, lineage, owners, and approvals
• Minimize, encrypt, restrict, and retain invocation payload logs deliberately
• Evaluate fairness, transparency, accountability, and documented limitations
• Design human review and incident escalation for high-risk outcomes

Optimize for price-to-performance after establishing a quality baseline. Operate semantic quality and business outcomes alongside normal service health.

• Track input/output tokens, requests, latency, errors, and throttles
• Prune irrelevant context and control response length
• Evaluate tiered model routing and model cascades
• Use caching only where freshness, safety, identity, and policy permit reuse
• Evaluate streaming, concurrency, batching, and throughput requirements
• Monitor vector-store query latency, relevance, freshness, and index health
• Use CloudWatch dashboards and alarms for operational and custom quality signals
• Use CloudTrail for API audit and X-Ray for cross-service tracing where applicable
• Configure Bedrock invocation logging only with approved privacy controls
• Build cost anomaly and per-workload attribution practices

Finish with repeatable evaluation and failure diagnosis. Practice explaining why an option satisfies all constraints—not merely which service name appears in the question.

• Version representative normal, difficult, ambiguous, and adversarial datasets
• Measure relevance, correctness, consistency, fluency, grounding, safety, latency, and cost
• Evaluate retrieval and generation separately for RAG
• Validate citation existence and support for generated claims
• Evaluate agent completion, tool choice, argument validity, loops, and unsafe actions
• Calibrate automated or model-based evaluation with human review
• Use regression gates, canaries, rollback, and synthetic workflows
• Troubleshoot context overflow, truncation, malformed requests, and prompt confusion
• Diagnose parsing, embedding, chunking, vectorization, filter, and reranking failures
• Complete timed original practice and review every distractor against AWS documentation