Is the AWS Bedrock Academy free?

Yes. The full course is free and includes 25 modules, 52 lessons, 52 labs, and quiz-based progression with no sign-up required.

How is Infinity AWS Bedrock Academy different from AWS Skill Builder?

Unlike AWS Skill Builder which often gates advanced hands-on labs behind paid subscriptions or AWS Builder ID logins, Infinity Tech Stack's AWS Bedrock Academy is 100% free, ungated, and browser-first with zero sign-up required. It includes 25 modules, 52 lessons, 52 hands-on labs, and 125+ quiz checks covering the Converse API, Bedrock Guardrails, AgentCore, and multi-model evaluations without requiring active AWS account credentials.

What is Amazon Bedrock and how does its architecture operate?

Amazon Bedrock is AWS's fully managed foundation model service providing serverless inference across frontier providers (Amazon Nova & Nova 2, Anthropic Claude 5.x, Meta Llama 4, DeepSeek-R1/V3, Mistral) via a unified Converse API, backed by AWS IAM, KMS encryption, and VPC PrivateLink isolation.

What is the Amazon Bedrock Converse API and why should you use it over InvokeModel?

The Converse API provides a unified request and response specification across foundation models on Bedrock. It standardizes system prompts, cachePoint prompt caching, guardrailConfig enforcement, structured tool use, and multi-turn message history without vendor-specific payload drift.

How does Amazon Bedrock AgentCore differ from Classic Bedrock Agents?

Classic Bedrock Agents entered maintenance mode (closed to new customers on July 30, 2026), succeeded by Amazon Bedrock AgentCore. AgentCore is AWS's modular, framework-agnostic platform (native to the open-source Strands Agents SDK, LangGraph, and CrewAI) featuring isolated microVM Runtime sessions, MCP Gateway tool endpoints with Cedar policy controls, episodic Memory, Identity & Consent Portal, Browser, Code Interpreter, and OpenTelemetry Observability.

How do Bedrock Guardrails enforce safety and PII redaction?

Bedrock Guardrails apply customizable safety policies across both user inputs and model outputs (or standalone via ApplyGuardrail). They filter multimodal text and image toxicity, enforce mathematical Automated Reasoning checks, score RAG Contextual Grounding, block prompt attacks, and mask or block sensitive PII.

What retrieval strategies are supported in Bedrock Knowledge Bases?

Bedrock Knowledge Bases support GraphRAG powered by Amazon Neptune Analytics for multi-hop reasoning, dense and HYBRID (vector + BM25) search across OpenSearch Serverless and Aurora pgvector, Amazon Kendra GenAI Index connectors, multimodal parsing via Bedrock Data Automation (BDA), and cross-encoder reranking with Cohere Rerank 3.5 and Amazon Rerank 1.0.

How does Bedrock Cross-Region Inference prevent 429 throttling exceptions?

Cross-region inference profiles (us., eu., apac., and global. prefixes) automatically route traffic across regional AWS capacity pools to eliminate localized throttling, with global. profiles offering ~10% cost savings. Newer frontier models (such as Claude 5.x, Nova Premier/Nova 2, Llama 4, and DeepSeek-R1) require inference profile IDs for on-demand invocation.

What are Amazon Nova models on Bedrock?

Amazon Nova is AWS's native foundation model family spanning Nova Micro, Nova Lite, Nova Pro, and Nova Premier, creative/voice models Nova Sonic (speech-to-speech), Nova Canvas (image), and Nova Reel (video), plus the 2026 Nova 2 reasoning and voice lineup (Nova 2 Lite, Nova 2 Pro, Nova 2.5 Sonic).

How does prompt caching operate on Amazon Bedrock?

Prompt caching uses cachePoint checkpoints in the Converse API (with a 5-minute default or optional 1-hour TTL) to keep static prefixes—system prompts, tool schemas, and reference documents—in memory, cutting cached input token costs by up to 90% and lowering latency by over 80%.

What is Provisioned Throughput (PT) and when should it be purchased?

Provisioned Throughput allocates dedicated model units to guarantee steady-state inference performance and strict SLA compliance for mission-critical production applications with predictable traffic volumes.

How does Bedrock model evaluation work for custom and foundation models?

Bedrock Model Evaluation supports automated algorithmic scoring across metrics like accuracy, robustness, and toxicity, as well as human-in-the-loop review teams via Amazon SageMaker Ground Truth integration.

How are Bedrock API calls secured with AWS KMS and PrivateLink?

Customer Managed Keys (CMKs) provide customer-controlled envelope encryption for model weights and Knowledge Base vector stores. AWS PrivateLink VPC endpoints keep inference traffic strictly within the customer's private virtual cloud.

How does Bedrock integrate with AWS Lambda and Amazon API Gateway?

Bedrock agents invoke AWS Lambda functions to execute enterprise actions, while Amazon API Gateway provides secure HTTP/REST endpoints with Cognito JWT authentication, usage tiers, and DDoS protection.

What telemetry and observability tools are available for Bedrock?

Amazon Bedrock streams runtime metrics (latency, token consumption, error codes) directly to Amazon CloudWatch, writes complete invocation audit records to CloudWatch Logs or S3, and integrates with AWS X-Ray for distributed agent tracing.

How can organizations implement chargeback and cost allocation for Bedrock?

Organizations apply AWS Cost Allocation Tags (e.g. CostCenter, Team, Environment) to inference profiles and application roles, analyzing usage via AWS Cost Explorer and establishing budget guardrails with AWS Budgets.

What chaos testing and disaster recovery patterns apply to Bedrock?

Resilience patterns include Route 53 Application Recovery Controller (ARC) regional failover, AWS Fault Injection Simulator (FIS) synthetic throttling injection, and circuit breakers with fallback response caches.

Can I earn an official certificate upon completing AWS Bedrock Academy?

Yes! Completing all 25 modules unlocks the official Infinity AWS Bedrock AI Architecture Mastery Certificate with your cryptographic verification ID and one-click LinkedIn badge sharing.

Escalation to stronger model tiers should be triggered by:

Policy and quality thresholds — Threshold-based escalation keeps cost controlled while protecting quality on hard tasks.

Choosing Models by Workload — Model Selection Strategy

← Back to Dashboard

1. Choosing Models by Workload2. A/B Testing and Rollout Gates

Choosing Models by Workload

📚 Model Selection Strategy⏱ 11 min⭐ 85 XP⌨ Hands-on lab

One Model Is Rarely Optimal

Bedrock hosts 100+ foundation models across Amazon, Anthropic, Meta, Mistral, DeepSeek, Moonshot AI (Kimi), MiniMax, Google DeepMind (Gemma), xAI (Grok), and OpenAI. Production systems route requests to different model tiers based on risk and required quality - paying frontier prices for trivial tasks is the most common AI budget leak.

A Practical Routing Matrix

Workload	Risk/complexity	Model tier	Why
Summarization, tagging, classification, voice turn-taking	Low	Amazon Nova Micro / Nova 2 Lite, Claude Haiku 5.5, Nova 2.5 Sonic (voice)	Sub-second latency, low cost, configurable light thinking
Customer-facing drafting, grounded RAG, UI generation	Medium	Amazon Nova Pro / Nova 2 Pro, Claude Sonnet 5.5, Claude Fable 5.1	Balanced reasoning, tool accuracy, and cost at volume
Complex reasoning, agentic coding, multi-step architecture	High	Claude Opus 5.5, Amazon Nova Premier, DeepSeek-R1	Frontier reasoning depth is the bottleneck, not token price
Open-weight / portability mandates	Varies	Meta Llama 4 (Scout 17B, Maverick 17B), DeepSeek-V3.2, Mistral Large, GPT-OSS	Open-weight governance and cross-cloud portability
Regulated decisions (credit, medical, legal)	Critical	Strongest tier + Automated Reasoning Guardrails + human gates	Mathematical policy proof and audit trails outweigh token cost

Routing Mechanics

Static routing - endpoint/task type maps to a model in config. Simple, predictable, covers most systems.
Dynamic routing - a cheap classifier scores request complexity, then dispatches. Adds cost/latency; justify with traffic diversity.
Escalation routing - try the cheap tier; if confidence/validation fails, retry on the premium tier. Great when most requests are easy.
Intelligent Prompt Routing (Bedrock feature) - serverless single-endpoint routing between models in the same family (e.g. Haiku ↔ Sonnet or Nova Lite ↔ Nova Pro) based on predicted prompt complexity, cutting spend by up to ~30%+ without sacrificing quality.

Keep the matrix versioned: the model space shifts fast - review your routing matrix quarterly with fresh eval scores. Last year's "best value" tier is often this year's overpriced one.

⌨ HANDS-ON LABBuild a Routing Matrix

⭐ +150 XP

Map support, analytics, coding, and compliance tasks to model families.

1Create a workload-to-model matrix file.

2Add an escalation rule.

lab-sandbox — simulated environment

INFINITY LAB SANDBOX v2.6 — simulated shell
Type the command for the current objective. Helpers: "hint", "solution", "clear".
$

OBJECTIVE 1 / 2 — type "hint" if stuck

🧪 Knowledge Check

Press 1-4 to select1 of 2

Why do mature teams use model routing?

Because one model is always best

To optimize quality-latency-cost across workloads

To avoid testing

To remove telemetry