AI Architecture
AI Architecture is the core of this site — concrete, diagram-driven reference architectures and design patterns for how AI-native systems actually get built, not checklists (see Architecture Playbooks) or foundational concepts (see Learning Hub). This section covers reference architectures for RAG platforms, multi-agent systems, AI gateways, and enterprise AI platforms — the system designs, trade-offs, and patterns that separate a working production architecture from a whiteboard diagram.
New to AI architecture? The five-part foundation series now lives in the Learning Hub — start there to build the fundamentals before diving into the reference architectures below.
Coming Soon
These reference architectures, patterns, and applied designs are in development and will publish on a rolling basis:
Reference Architectures
End-to-end system designs, with diagrams, for how AI-native platforms actually get built.
Reference Architecture: Production RAG Platform
A full reference architecture for retrieval-augmented generation at production scale — ingestion, chunking, vector store, retrieval, and generation layers, with the trade-offs at each hop.
Reference Architecture: Multi-Agent Customer Support System
How a supervisor-worker multi-agent system is actually wired together — orchestration, tool access, escalation paths, and where human-in-the-loop fits.
Reference Architecture: Enterprise AI Gateway Deployment
A reference deployment for a production AI gateway — routing, auth, rate limiting, observability, and multi-provider failover.
Edge & On-Device AI
Reference architectures for running AI outside the cloud — smaller models, on-prem inference, and hybrid escalation patterns.
Reference Architecture: SLM Deployment at the Edge
Deploying small language models on-prem or on-device — model size, latency, and cost trade-offs versus a cloud LLM call.
Cloud vs. Edge Inference: Hybrid Escalation Patterns
A reference pattern for running a small model on-device and escalating to the cloud only when it's actually needed.
Patterns & Trade-offs
Named, reusable architecture patterns and head-to-head design trade-offs.
Pattern: Sidecar Guardrails for LLM Applications
A named pattern for decoupling safety and policy enforcement from application logic using a sidecar.
Hub-and-Spoke vs. Mesh for Multi-Agent Orchestration
Two competing topologies for coordinating multiple agents, and when each one actually holds up.
Centralized vs. Federated AI Platform Architecture
The real trade-offs between one shared AI platform and federated, team-owned platforms — cost, governance, and speed.
Applied Design
Worked examples that show real architectural judgment, not just theory.
How I'd Architect It: An AI-Native Retail Platform
A worked example — how I'd design the architecture for an AI-native retail platform from the ground up.
Bridging Legacy Enterprise Architecture and AI Systems
How AI systems actually plug into existing enterprise architecture — APIs, data platforms, and governance, without a rip-and-replace.