AI Architecture Patterns Skill

Purpose

Provide expert guidance on selecting and implementing enterprise AI architecture patterns for production systems. This skill contains battle-tested patterns from real-world deployments and the AI Architect Academy.

When to Use

• Designing new AI systems
• Evaluating architecture options
• Selecting patterns for specific use cases
• Understanding tradeoffs between approaches
• Getting implementation guidance

Core Patterns Library

1. AI Gateway Pattern

Problem: Multiple AI services with inconsistent interfaces, no centralized security, and limited observability create management complexity and security risks.

Solution: Deploy a centralized AI gateway that provides unified authentication, rate limiting, request/response logging, and model routing for all AI services.

Key Components:

• API Gateway (Kong, AWS API Gateway, OCI API Gateway)
• Authentication Service (OAuth2, API Keys)
• Rate Limiter (Redis-based)
• Request Logger (OpenTelemetry)
• Model Router

When to Use:

• Multiple AI providers in your stack
• Need centralized security controls
• Want unified logging and monitoring
• Cost allocation across teams

When NOT to Use:

• Single AI provider with simple use case
• Ultra-low latency requirements (<10ms)
• Early prototyping

---

2. RAG Production Pattern

Problem: LLMs hallucinate and lack access to enterprise-specific knowledge, making them unreliable for business-critical applications.

Solution: Implement a RAG pipeline with document ingestion, chunking, embedding, vector storage, retrieval, and augmented generation with source citations.

Key Components:

• Document Ingestion Pipeline
• Text Chunking Service (semantic, fixed, hybrid)
• Embedding Model (OpenAI, Cohere, Local)
• Vector Database (Pinecone, Weaviate, pgvector)
• Retrieval Service with Reranking
• LLM with RAG prompt template

When to Use:

• Customer support knowledge base
• Internal document Q&A
• Legal/compliance document analysis
• Technical documentation assistants

When NOT to Use:

• General creative writing
• Real-time frequently changing data
• Very small document corpus (<100 docs)

Implementation Tips:

• Start with fixed-size chunks (512-1024 tokens)
• Add metadata extraction for filtering
• Implement hybrid search (keyword + semantic)
• Use reranking for improved precision

---

3. Multi-Agent Orchestration Pattern

Problem: Complex tasks require multiple specialized capabilities that exceed what a single LLM prompt can handle reliably.

Solution: Decompose complex workflows into specialized agents with an orchestrator that coordinates task distribution, handoffs, and result aggregation.

Key Components:

• Orchestrator Agent (workflow coordinator)
• Specialized Worker Agents (domain experts)
• Task Queue (for async processing)
• State Management (context preservation)
• Handoff Protocol (agent-to-agent communication)
• Result Aggregator

When to Use:

• Complex workflows with 5+ distinct steps
• Tasks requiring different expertise
• Autonomous systems
• Workflows with branching logic

When NOT to Use:

• Simple single-step tasks
• When cost is primary constraint
• High-volume, low-complexity operations

Frameworks:

• LangGraph (graph-based orchestration)
• Claude Agent SDK
• AutoGen / CrewAI

---

4. MCP Server Architecture

Problem: N agents x M tools = N*M integrations. Each AI agent needs custom code to integrate with each tool.

Solution: Implement MCP (Model Context Protocol) servers that provide standardized interfaces for tools, resources, and prompts.

Key Components:

• MCP Server (Node.js or Python)
• Tool Definitions (JSON Schema)
• Resource Providers
• Prompt Templates
• Transport Layer (stdio, SSE)

When to Use:

• Building tools for multiple AI agents
• Creating reusable integrations
• Claude Code environments
• Enterprise tool standardization

Implementation:

import { Server } from '@modelcontextprotocol/sdk/server';

const server = new Server({
  name: 'my-mcp-server',
  version: '1.0.0'
});

server.tool('search', {
  description: 'Search documents',
  inputSchema: {
    type: 'object',
    properties: {
      query: { type: 'string' }
    }
  },
  handler: async ({ query }) => {
    // Implementation
  }
});

---

5. LLMOps Pipeline Pattern

Problem: LLM applications lack mature DevOps practices, leading to unpredictable quality and difficult rollbacks.

Solution: Implement prompt versioning, automated evaluation, staged deployments, and continuous monitoring.

Key Components:

• Prompt Version Control (Git, Promptfoo)
• Evaluation Dataset
• Automated Eval Pipeline
• Deployment Orchestrator
• Monitoring Dashboard
• Rollback Mechanism

When to Use:

• Production LLM applications
• Teams with multiple prompt engineers
• Regulated industries
• High-stakes AI applications

Evaluation Metrics:

• Accuracy (vs golden answers)
• Latency (p50, p95, p99)
• Cost per request
• User satisfaction scores

---

6. Vector Database Selection Framework

Problem: Many vector database options with different tradeoffs. Wrong choice leads to expensive migrations.

Solution: Structured decision framework evaluating scale, features, operations, and cost.

Selection Matrix:

Scale	Recommendation
<1M vectors	pgvector (simple), Chroma (prototyping)
1-100M vectors	Weaviate, Qdrant (self-hosted)
100M+ vectors	Pinecone, Milvus (managed)

Key Considerations:

• Hybrid search support
• Metadata filtering
• Multi-tenancy
• Backup/restore
• Managed vs self-hosted

---

7. AI Center of Excellence Framework

Problem: Scattered AI initiatives across organization lead to duplicated effort, inconsistent quality, and security gaps.

Solution: Establish centralized governance with standardized patterns, reusable components, and shared infrastructure.

Key Components:

• Pattern Library (this skill!)
• Governance Framework
• Shared Infrastructure
• Training Program
• Review Board
• Metrics Dashboard

Governance Areas:

• Model selection criteria
• Security standards
• Cost controls
• Ethical guidelines
• Incident response

---

8. Security & Governance Pattern

Problem: AI introduces new security vectors: prompt injection, data leakage, model manipulation.

Solution: Implement AI-specific security controls including guardrails, PII handling, and audit logging.

Key Controls:

• Input Guardrails (prompt injection detection)
• Output Guardrails (content filtering)
• PII Detection & Redaction
• Audit Logging
• Access Control
• Compliance Reporting

Guardrails Implementation:

from guardrails import Guard

guard = Guard.from_pydantic(output_class=SafeResponse)

response = guard(
    llm.invoke,
    prompt=user_input,
    on_fail="reask"
)

---

Pattern Selection Decision Tree

START: What type of AI system?
│
├── Document/Knowledge Q&A
│   └── → RAG Production Pattern
│       ├── Need multiple models? → + AI Gateway
│       └── Sensitive data? → + Security & Governance
│
├── Autonomous Agents
│   └── → Multi-Agent Orchestration
│       ├── Many tools? → + MCP Servers
│       └── Production deployment? → + LLMOps
│
├── Enterprise AI Platform
│   └── → AI Gateway + AI CoE Framework
│       ├── Cost concerns? → + Cost Optimization
│       └── Compliance? → + Security & Governance
│
└── Content Generation
    └── → AI Gateway + LLMOps
        └── Quality critical? → + Evaluation Pipeline

Pattern Combinations Matrix

Use Case	Primary	Secondary	Tertiary
Customer Support Bot	RAG	AI Gateway	Security
Code Assistant	Multi-Agent	MCP Servers	LLMOps
Document Intelligence	RAG	Vector DB	AI Gateway
Enterprise AI Platform	AI Gateway	AI CoE	Security
Research Assistant	RAG	Multi-Agent	LLMOps

Cloud Provider Mapping

AWS

• AI Gateway: API Gateway + Lambda
• RAG: Bedrock + OpenSearch
• Vector DB: OpenSearch, Aurora pgvector

GCP

• AI Gateway: Cloud Endpoints + Cloud Functions
• RAG: Vertex AI + Matching Engine
• Vector DB: Matching Engine, AlloyDB

Azure

• AI Gateway: API Management + Functions
• RAG: Azure OpenAI + AI Search
• Vector DB: AI Search, Cosmos DB

OCI

• AI Gateway: API Gateway + Functions
• RAG: OCI GenAI + OpenSearch
• Vector DB: OpenSearch, PostgreSQL

Resources

• GitHub: https://github.com/frankxai/ai-architect-academy
• Patterns Library: /01-design-patterns
• Learning Paths: /02-learning-paths
• Templates: /AI CoE Templates

Related Skills

• mcp-architecture - MCP server development
• claude-sdk - Agent development with Claude
• langgraph-patterns - Graph-based agent workflows
• oci-services-expert - Oracle Cloud guidance

---

Part of the AI Architect Academy by FrankX.AI