genai-architecture-guide

# GenAI Architecture Guide for FDE & Applied AI Interviews Learn how to design and explain production GenAI systems—from model APIs and RAG to agents, evaluation, security, and operations. This guide is written for **Forward Deployed Engineer (FDE)**, **Applied AI Engineer**, and **AI Engineer** candidates who need to reason about customer-facing AI systems, not memorize a fashionable framework. **[Start the complete FDE GenAI course →](https://www.fdehandbook.com/learn/genai-architecture/overview?utm_source=github&utm_medium=referral&utm_campaign=genai_field_guide&utm_content=readme_top)**  The map separates the application control plane from retrieval, model runtime, and external actions. Evaluation, observability, guardrails, security, and release engineering span the entire system rather than appearing as afterthoughts. ## What makes this guide different - **Architecture decisions, not framework tutorials.** Learn when to use a direct model call, RAG, a deterministic workflow, or an agent. - **Production trade-offs, not toy chatbots.** Treat permissions, evaluation, latency, cost, observability, and recovery as part of the design. - **Structured for system-design interviews.** Build the vocabulary and decision process expected in FDE and Applied AI interviews. - **Connected to deliberate practice.** Continue into guided architecture scenarios, coding exercises, and an AI mock interviewer when you are ready to rehearse. ## Start here 1. [Why GenAI changes software architecture](01-foundations-of-genai-systems/1.1-from-traditional-software-to-genai-systems.md) 2. [What makes a system agentic?](01-foundations-of-genai-systems/1.3-what-makes-a-system-agentic.md) 3. [Model + Harness](01-foundations-of-genai-systems/1.4-model-plus-harness.md) 4. [RAG Architecture Pattern](03-genai-architecture-patterns/3.3-rag-architecture-pattern.md) 5. [Why AI Reliability Is Different](04-reliable-ai-systems/4.1-why-ai-reliability-is-different.md) ## Architecture pattern selection cheat sheet Start with the simplest pattern that satisfies the requirement. Add autonomy only when the system genuinely needs to choose its path at runtime. | Requirement | Start with | Add only when needed | |---|---|---| | Generate, classify, extract, or transform bounded input | Direct model call | Structured output and deterministic validation | | Maintain a conversational interface | Chatbot | Explicit history selection and state | | Answer from private, current, or citable knowledge | RAG | Hybrid retrieval, reranking, and access-aware filtering | | Execute a known multi-step business process | Deterministic workflow | Model steps inside application-controlled transitions | | Choose tools or next steps dynamically | Agentic workflow | Budgets, stop conditions, scoped permissions, and recovery | | Perform an irreversible or high-risk action | Human approval | Audit evidence and explicit authorization before execution | Before choosing a pattern, ask: - Does the model already have the required information? - Are the execution steps known in advance? - Must the system take an external action? - What is the cost of a plausible but wrong result? - How will we evaluate, trace, recover, and stop it? ## Course contents | Chapter | What it covers | |---|---| | [01 · Foundations of GenAI Systems](01-foundations-of-genai-systems/) | Six lessons, from traditional software to the new role of the engineer | | [02 · Building Blocks of GenAI Systems](02-building-blocks-of-genai-systems/) | Models, context, retrieval, tools, MCP, state, memory, and permissions | | [03 · GenAI Architecture Patterns](03-genai-architecture-patterns/) | Direct LLM, RAG, workflows, agents, skills, interoperability, and human review | | [04 · Reliable AI Systems](04-reliable-ai-systems/) | Evaluation, security, guardrails, observability, governance, and readiness | | [05 · Production-Grade GenAI Engineering](05-production-grade-genai-engineering/) | Specifications, production architecture, deployment, data, CI/CD, scale, and operations | Every website subsection is stored as its own numbered Markdown file with the same lesson title. ## Open guide vs. complete handbook This repository explains the core mental models and production principles. The complete FDE Handbook adds deeper course chapters, guided walkthroughs, runnable coding exercises, closed-book practice, progress tracking, and an AI mock interviewer. Locked lessons use the same **Continue the course** or **Member chapter** boundary shown in the Handbook. ## Continue with practice - [GenAI architecture scenarios](https://www.fdehandbook.com/practice/genai-architecture?utm_source=github&utm_medium=referral&utm_campaign=genai_field_guide&utm_content=readme_practice) - [Build an agent loop from scratch](https://www.fdehandbook.com/practice/coding/agent-loop-from-scratch?utm_source=github&utm_medium=referral&utm_campaign=genai_field_guide&utm_content=readme_agent_loop) - [AI Mock Interviewer](https://www.fdehandbook.com/practice/mock/ai-interviewer?utm_source=github&utm_medium=referral&utm_campaign=genai_field_guide&utm_content=readme_mock_interview) - [30-day FDE study plan](https://www.fdehandbook.com/learn/study-plan?utm_source=github&utm_medium=referral&utm_campaign=genai_field_guide&utm_content=readme_study_plan) The previous question-bank README is preserved in [`Forward-Deployed Engineer (FDE) Interview Prep/question-bank.md`](<Forward-Deployed Engineer (FDE) Interview Prep/question-bank.md>). ## License Course content is available under [CC BY-NC 4.0](LICENSE). You may share and adapt it with attribution for non-commercial use. The synchronization script is available under the [MIT License](scripts/LICENSE). Built for engineers preparing for Forward Deployed Engineer, Applied AI Engineer, Solutions Engineer, Customer Engineer, and Field Engineer roles.