Skills — AWS AIP-C01 Domain 2 Inventory

This document maps all 25 skills from AWS Certified AI Practitioner (AIP-C01) Domain 2 to their MangaAssist implementation. Domain 2 covers the design, deployment, and operation of agentic AI solutions in a production AWS environment.

Source material: Implementation-Integration-Domain2/

Domain 2 Task Map

Task 2.1  Agentic AI Solutions          (7 skills)
Task 2.2  Model Deployment              (3 skills)
Task 2.3  Enterprise Integration        (5 skills)
Task 2.4  FM API Integrations           (4 skills)
Task 2.5  Application Integration       (6 skills)

Task 2.1 — Agentic AI Solutions

Skill 2.1.1 — Intelligent Autonomous Systems

Design agents that plan, execute multi-step tasks, and adapt based on observations.

Core concepts: Strands Agents, AWS Agent Squad, Model Context Protocol (MCP), memory and state management, agent lifecycle (Initialize → Plan → Act → Observe → Reflect)

MangaAssist implementation: - Orchestrator agent uses Strands Agents on Bedrock (Claude 3.5 Sonnet) - AWS Agent Squad coordinates four specialised sub-agents (ProductSearch, OrderStatus, Recommendation, MangaQA) - Session state persisted in DynamoDB; short-term context in ElastiCache Redis - MCP protocol connects orchestrator to seven domain-specific tool servers

Key files: agents.md, RAG-MCP-Integration/08-mcp-orchestration-router.md

Skill 2.1.2 — Advanced Problem-Solving Systems

Build agents that use structured reasoning to solve complex, multi-hop problems.

Core concepts: ReAct (Reason + Act) loop, chain-of-thought prompting, AWS Step Functions for durable multi-step workflows

MangaAssist implementation: - Orchestrator implicitly executes ReAct: reasons over tool manifests, acts by calling tools, observes results, and reflects before generating the final answer - Chain-of-thought is encouraged via system prompt for recommendation explanations - Step Functions used for long-running workflows (return processing, escalation) that outlive a single HTTP request

Key files: LLD-Questions/topic-deep-dives/01-orchestrator-request-flow.md

Skill 2.1.3 — Safeguarded AI Workflows

Implement controls that prevent runaway execution, enforce safety boundaries, and protect users.

Core concepts: Stopping conditions, iteration limits, wall-clock timeouts, circuit breakers, guardrails (input and output)

MangaAssist implementation: - Max 10 tool-call iterations per user message; 8-second wall-clock timeout on the agentic loop - Input guardrail: PII detection and redaction before tool dispatch - Output guardrail: ASIN validation, price sanity check, link validation before streaming - Amazon Bedrock Guardrails for toxicity and harmful content filtering - Circuit breakers on each MCP server (5 failures / 60 s → open circuit)

Key files: LLD-Questions/topic-deep-dives/05-guardrails-validation-safety.md, AI-Safety-Security-Governance/01-input-output-safety-controls/

Skill 2.1.4 — Model Coordination Systems

Orchestrate multiple models to improve accuracy, coverage, or latency through ensembles and routing.

Core concepts: Multi-model routing, ensemble strategies, model-as-judge, speculative decoding

MangaAssist implementation: - Two-stage intent classifier: Stage 1 is rule-based regex (fast, cheap); Stage 2 is fine-tuned DistilBERT on SageMaker (≥ 0.6 confidence threshold) - Orchestrator (Claude 3.5 Sonnet) handles open-ended generation; DistilBERT handles classification — different models for different cost/latency profiles - Cross-encoder reranker (BGE-reranker-v2-m3) runs as a third model in the RAG pipeline

Key files: LLD-Questions/topic-deep-dives/02-intent-classification-entity-resolution.md

Skill 2.1.5 — Collaborative AI Systems

Design workflows where humans and AI collaborate, including review checkpoints and escalation paths.

Core concepts: Human-in-the-loop, approval gates, review workflows, escalation routing, Amazon Connect integration

MangaAssist implementation: - Escalation triggered on: confidence < 0.6, explicit user request, guardrail violation, > 2 consecutive tool failures - On escalation: session context snapshot serialised to DynamoDB → SNS event → Amazon Connect agent queue - Human agent receives full conversation summary; session resumes without user re-explaining context

Key files: mangaassist_workflow_interview_pack/01_critical_workflows_interview_grill.md

Skill 2.1.6 — Intelligent Tool Integrations

Define, register, and robustly call external tools from within an agent loop.

Core concepts: Tool definition (name, description, JSON schema), error handling, retry policies, idempotency, tool result validation

MangaAssist implementation: - All seven MCP servers expose typed tool schemas; schemas are injected into the orchestrator's system prompt - Tool descriptions are the primary routing mechanism — description quality drives selection accuracy - Retry: up to 2 retries with exponential backoff (100 ms → 200 ms → 400 ms); idempotent by design (read-only tools) - Tool results validated (type, required fields) before insertion into context window

Key files: RAG-MCP-Integration/10-mcp-basics-and-amazon-backend-integration.md, subagents.md

Skill 2.1.7 — Model Extension Frameworks

Deploy and manage MCP servers that extend model capabilities with custom domain logic.

Core concepts: MCP protocol mechanics, Lambda vs ECS Fargate MCP servers, access patterns, versioning

MangaAssist implementation: - Seven MCP servers, all deployed as ECS Fargate tasks (not Lambda) due to connection persistence requirements - Each server registered in AWS Cloud Map; orchestrator resolves endpoints at startup - MCP server versioning: breaking schema changes require a new tool name (backward-compatible evolution preferred) - Access pattern: orchestrator → internal ALB → ECS Fargate task → downstream AWS data services

Key files: ECS-Fargate-Lambda/, Implementation-Integration-Domain2/Skill-2.1.7-Model-Extension-Frameworks/

Task 2.2 — Model Deployment Strategies

Skill 2.2.1 — Deploy FMs Based on Application Needs and Performance Requirements

Choose the right deployment pattern (Lambda, Bedrock Provisioned Throughput, SageMaker endpoint) to match latency, throughput, and cost requirements per use case.

MangaAssist implementation: - Claude 3.5 Sonnet on Bedrock on-demand for interactive chat (spiky traffic, no reserved capacity needed) - DistilBERT intent classifier on SageMaker real-time endpoint (ml.g4dn.xlarge) for consistent 500 QPS classification - Cross-encoder reranker (BGE-reranker-v2-m3) on SageMaker with auto-scaling - Amazon Titan Embeddings V2 called via Bedrock API for document indexing

Skill 2.2.2 — Deploy FM Solutions Addressing Unique LLM Challenges

Handle LLM-specific deployment concerns: long context windows, streaming tokens, cold starts, context length limits, and hallucination at the deployment boundary.

MangaAssist implementation: - Streaming via InvokeModelWithResponseStream to eliminate cold-start perceived latency - Prompt caching for system prompts and tool manifests (target cache hit rate > 85 %) - Context window budget enforced at the orchestrator layer (hard cap on conversation history growth) - Temperature 0.3 for grounded factual responses; post-generation ASIN and price validation catches hallucinated product details

Skill 2.2.3 — Develop Optimized FM Deployment Approaches

Apply quantisation, distillation, and inference optimisation to reduce cost and latency.

MangaAssist implementation: - Rule-based Stage 1 classifier runs in-process (zero extra cost, zero latency overhead) - QLoRA (4-bit quantisation) applied during fine-tuning of the intent classifier - Knowledge distillation used to compress the cross-encoder reranker for lower SageMaker instance cost - Batch inference via Bedrock batch API for nightly review summarisation (not latency-sensitive)

Key files: Fine-Tuning-Foundational-Models/Model-Compression-Optimization/

Task 2.3 — Enterprise Integration

Skill 2.3.1 — Create Enterprise Connectivity Solutions

Integrate FM capabilities into existing enterprise architectures — legacy systems, databases, and third-party APIs — without disrupting current workflows.

MangaAssist implementation: - RDS PostgreSQL (existing order management system) accessed by the Order & Inventory MCP via a read replica; adapter pattern isolates the MCP server from direct schema coupling - EventBridge + SQS used for loose coupling between Amazon Connect escalation events and the MangaAssist session system

Skill 2.3.2 — Develop Integrated AI Capabilities for Existing Applications

Embed GenAI features into existing applications via microservice patterns, webhooks, and event-driven integration.

MangaAssist implementation: - Chat widget embedded in the existing Amazon JP Manga product page via iframe + postMessage API — no changes to the host application - Kinesis Data Streams: click/purchase events → Lambda consumer → trending aggregator (5-second window) - DynamoDB Streams: inventory deltas → Lambda → ElastiCache Redis cache invalidation - SNS: escalation events fan out to Amazon Connect and the analytics pipeline simultaneously

Skill 2.3.3 — Create Secure Access Frameworks

Apply IAM roles, VPC isolation, identity federation, and RBAC to GenAI workloads.

MangaAssist implementation: - Each MCP server has its own IAM execution role scoped to its specific data sources (least privilege) - All inter-service communication within VPC; no MCP server is publicly addressable - Cognito user pools for customer authentication; JWT validated before any order-related tool call - At-rest encryption via KMS on DynamoDB, S3, RDS, ElastiCache; TLS 1.2+ in transit

Key files: AI-Safety-Security-Governance/02-data-security-privacy-controls/, 12-security-privacy.md

Skill 2.3.4 — Develop Cross-Environment AI Solutions

Maintain environment parity (dev/staging/prod) and handle cross-region deployments for GenAI workloads.

MangaAssist implementation: - CDK stacks parameterised by environment; staging uses spot instances and smaller SageMaker instance types - Feature flags in AWS AppConfig control model version rollout (canary 5 % → full promotion) per environment - Bedrock cross-region inference profiles: primary us-east-1, failover us-west-2

Skill 2.3.5 — Implement CI/CD Pipelines and GenAI Gateway Architectures

Automate model deployment, RAG knowledge base refresh, and infrastructure changes via dedicated pipelines.

MangaAssist implementation: - Separate pipelines for: application code, IaC (CDK), ML model deployment, RAG knowledge base refresh, monitoring config - Knowledge base pipeline: policy doc change in S3 → EventBridge → Step Functions → re-chunk → re-embed → OpenSearch index swap - SageMaker Model Registry → canary rollout → CloudWatch alarm gate → auto-rollback on regression

Key files: CI-CD-Pipeline-User-Stories/

Task 2.4 — FM API Integrations

Skill 2.4.1 — Create Flexible Model Interaction Systems

Design FM interaction layers that support both synchronous (real-time) and asynchronous (background/batch) invocation patterns.

MangaAssist implementation: - Bedrock InvokeModel (synchronous) for interactive chat — P99 < 3 seconds - Bedrock batch API for nightly offline review summarisation (non-latency-sensitive) - Async invocation via SQS queue for knowledge base embedding refresh - Step Functions async task token pattern for return processing workflows (outlives a single HTTP request) - Connection pooling via Bedrock client reuse in the orchestrator process

Skill 2.4.2 — Develop Real-Time AI Interaction Systems (Streaming)

Stream model output tokens to clients to reduce perceived latency and improve interactive experience.

MangaAssist implementation: - Bedrock InvokeModelWithResponseStream used for all user-facing generation - WebSocket connection (API Gateway) carries the token stream from orchestrator to frontend - First token target: < 1 second; client renders incrementally - Streaming reconnect logic handles dropped WebSocket connections without losing session state

Key files: LLD-Questions/topic-deep-dives/06-api-contracts-streaming-and-escalation.md

Skill 2.4.3 — Create Resilient FM Systems

Build retry logic, circuit breakers, fallback models, and observability to maintain availability when FM APIs degrade.

MangaAssist implementation: - Exponential backoff on Bedrock throttling (429) with jitter; max 3 retries - Circuit breakers on each MCP server (5 failures / 60 s → open circuit) - CloudWatch alarms on Bedrock latency and error rate; auto-triggers cross-region failover - Rate limiting enforced at API Gateway layer to protect downstream Bedrock quotas

Skill 2.4.4 — Develop Intelligent Model Routing Systems

Route requests dynamically across models based on cost, capability, latency, or load.

MangaAssist implementation: - Bedrock cross-region inference profiles: primary us-east-1, failover us-west-2 - Fallback to Claude 3 Haiku for non-complex queries during high-load periods (cost + speed) - Two-stage intent classifier routes requests: rule-based (fast, free) → DistilBERT (medium cost) → Claude full generation (highest cost) — only escalates when confidence is insufficient at each stage

Task 2.5 — Application Integration Patterns

Skill 2.5.1 — Create FM API Interfaces for GenAI Workloads

Build API layers that handle streaming responses, enforce token limits, and manage timeouts specific to FM workloads — shielding downstream consumers from FM-specific complexity.

MangaAssist implementation: - REST API (API Gateway) for session lifecycle (create, close, get history) - WebSocket API (API Gateway) for real-time token streaming (no 29 s REST timeout constraint) - Token budget validation at the authoriser layer; input truncation enforced before Bedrock invocation - Versioned with v1/ prefix; breaking changes require a new version

Key files: API-Design-and-Testing/

Skill 2.5.2 — Develop Accessible AI Interfaces

Build low-code, embedded, and accessible UI components that surface AI capabilities to non-technical users and business teams.

MangaAssist implementation: - Chat widget embedded in Amazon JP Manga product pages via iframe + postMessage API (no host-page code changes) - Streaming tokens rendered progressively; typing indicator shown while waiting for first token - Escalation UI: seamless handoff to Amazon Connect CCP embedded widget — user never leaves the product page

Key files: 05-website-integration.md

Skill 2.5.3 — Create Business System Enhancements

Augment existing business systems — CRM, order management, analytics — with AI-generated insights and actions.

MangaAssist implementation: - Return initiation by MangaAssist writes directly to order management RDS via Order & Inventory MCP - Session close event publishes to Kinesis → Redshift for analytics (intent distribution, resolution rate, escalation rate) - Escalation events published to SNS → Amazon Connect CRM case creation

Skill 2.5.4 — Enhance Developer Productivity

Provide internal SDKs, AI-assisted development tools, and testing infrastructure to accelerate team adoption of GenAI capabilities.

MangaAssist implementation: - Internal MCP SDK (thin wrapper over the protocol) enabling any team to register a new tool server - Integration test harness in CI/CD validates all tool schemas against live endpoints on every deploy - AI-assisted prompt evaluation tooling: automated regression tests compare new prompt versions against golden answer sets

Skill 2.5.5 — Develop Advanced GenAI Applications

Apply RAG, prompt chaining, caching, and hybrid retrieval architectures to build production-grade AI applications.

MangaAssist implementation: - Full RAG pipeline: embed (Titan V2) → hybrid retrieve (BM25 0.3 + dense 0.7) → metadata filter → cross-encoder rerank → grounded generation - Prompt caching for system prompts and tool manifests (85 %+ cache hit target) - Two-tier caching: Redis for hot session state, DynamoDB for durable session history - Prompt chaining: entity extraction → tool dispatch → grounded synthesis (three distinct prompt stages per turn)

Key files: RAG-MCP-Integration/09-rag-retrieval-pipeline-deep-dive.md, LLD-Questions/topic-deep-dives/04-rag-indexing-retrieval-re-ranking.md

Skill 2.5.6 — Improve Troubleshooting Efficiency for FM Applications

Build observability, structured logging, and debugging frameworks to quickly diagnose failures across the LLM, retrieval, and orchestration layers.

MangaAssist implementation: - X-Ray distributed tracing: root trace per user request spans orchestrator → MCP servers → data stores - Structured logging with correlation IDs propagated across all services - Four failure taxonomies: retrieval failures (low recall), generation failures (hallucination, refusal), tool failures (timeout, schema mismatch), system failures (infrastructure) - CloudWatch dashboards: intent accuracy, tool call success rate, RAG recall@5, escalation rate, P99 latency

Key files: Monitoring-GenAI-Systems/, LLD-Questions/topic-deep-dives/07-analytics-observability-and-feedback-loops.md, Troubleshoot-GenAI-Applications/

Skill Cross-Reference

Related Documents

• agents.md — Orchestrator agent detail
• subagents.md — MCP server sub-agents
• Implementation-Integration-Domain2/ — Full skill documentation
• 10-ai-llm-design.md — LLM and intent classification design
• 11-scalability-reliability.md — Scaling strategies