Lilly Tech Systems
AI Design Patterns
Master the proven, reusable architectural patterns that power every production AI system — from RAG and Agents to Prompt Chaining, Guardrails, Ensemble Voting, and Event-Driven AI. Learn when to apply each pattern, how they combine, and how to avoid the most common anti-patterns in AI system design.
Your Learning Path
Follow these lessons in order to build a complete mental model of AI design patterns, or jump to any pattern that your project needs right now.
1. Introduction
What are AI design patterns, why they matter, the five pattern categories (Data, Inference, Orchestration, Safety, Optimization), and a complete overview of all 12 patterns covered in this course.
2. RAG Pattern
Retrieval-Augmented Generation: naive vs advanced vs modular RAG, chunking strategies, dense/sparse/hybrid retrieval, reranking, query transformation, self-RAG, and complete code examples.
3. Agent & ReAct
AI agents with the observe-think-act loop, ReAct (Reason + Act), tool use, planning patterns, multi-agent architectures (supervisor, debate, swarm), and agent memory systems.
4. Prompt Chaining
Break complex tasks into sequential LLM calls. Chain-of-thought decomposition, conditional branching, validation gates, map-reduce for documents, and error recovery strategies.
5. Router & Gateway
Intelligent request routing across models and providers. Semantic routers, cost-based routing, latency-aware load balancing, API gateway patterns, and model fallback chains.
6. Cascade & Fallback
Escalate from cheap/fast models to expensive/powerful ones only when needed. Confidence-based cascading, graceful degradation, circuit breakers, and cost optimization strategies.
7. Ensemble & Voting
Combine outputs from multiple models for higher accuracy. Majority voting, weighted ensembles, model-as-judge, debate protocols, and when ensembles justify the extra cost.
8. Human-in-the-Loop
When and how to involve humans in AI pipelines. Approval workflows, active learning, confidence thresholds, feedback loops, and designing effective human review interfaces.
9. Guardrails & Safety
Input validation, output filtering, PII detection, prompt injection defense, content moderation, structured output enforcement, and building layered safety systems.
10. Cache & Optimization
Semantic caching, exact-match caching, prompt compression, KV-cache reuse, batching strategies, and reducing LLM costs by 10x without sacrificing quality.
11. Fan-out / Fan-in
Parallel AI execution for throughput and quality. Scatter-gather, map-reduce with LLMs, parallel tool calls, aggregation strategies, and handling partial failures.
12. Event-Driven AI
Reactive AI systems triggered by events. Webhook-driven inference, streaming pipelines, async processing, event sourcing for AI state, and real-time AI with message queues.
13. Pattern Selection Guide
The decision framework: given a problem, which pattern(s) to apply. Decision trees, pattern compatibility matrix, real-world case studies, and a step-by-step selection checklist.
What You'll Learn
By the end of this course, you will be able to:
Recognize Core Patterns
Identify the 12 most important AI design patterns used in production systems — RAG, Agents, Prompt Chaining, Routers, Cascades, Ensembles, Guardrails, and more — and understand the problem each one solves.
Implement Patterns in Code
Build each pattern from scratch with Python, LangChain, and direct API calls to Claude, OpenAI, and open-source models. Every lesson includes working code you can adapt to your own projects.
Combine Patterns Effectively
Compose multiple patterns into robust AI architectures. Learn which patterns complement each other (RAG + Guardrails + Cache) and which combinations create unnecessary complexity.
Avoid Common Anti-Patterns
Recognize and prevent the most expensive mistakes in AI system design — over-engineering, missing guardrails, ignoring caching, single points of failure, and agent infinite loops.
Who This Course Is For
This course is designed for developers and architects who are already building or planning to build AI-powered applications.
Backend Developers
You know how to build APIs and services. Now learn the patterns specific to AI — how to structure LLM calls, handle non-deterministic outputs, and build reliable AI pipelines.
AI Engineers
You have built prototypes with LLMs. Now learn the production patterns that separate demos from reliable systems — caching, fallbacks, guardrails, and observability.
Solution Architects
You design systems at scale. Learn the AI-specific patterns you need to add to your architecture toolbox — alongside the familiar patterns you already know from distributed systems.
Prerequisites
To get the most from this course, you should have:
- Python basics: Variables, functions, classes, async/await. You do not need to be an expert, but you should be comfortable reading Python code.
- API experience: Familiarity with REST APIs, HTTP requests, and JSON. Most AI patterns involve calling model APIs.
- LLM exposure: You have used ChatGPT, Claude, or a similar LLM at least casually. You understand what a prompt is and what an LLM response looks like.
- No ML math required: This course focuses on architecture and engineering patterns, not on training models or linear algebra. You do not need to understand backpropagation or gradient descent.