Chapter 1: Introduction to Building AI Applications

+-------------------------------------------------------------------------------+
|                            THE MODERN AI APP STACK                            |
+-------------------------------------------------------------------------------+
|  Application Layer     | UI, Chat interfaces, Copilots, API Gateway          |
|  Orchestration Layer   | RAG pipelines, Agent frameworks (LangGraph, ReAct)   |
|  Context Layer         | Vector DBs, Hybrid Search, Enterprise Knowledge Base  |
|  Foundation Model Layer| Proprietary APIs (GPT-4o, Claude) & Open Weights     |
|  Hardware & Serving    | vLLM, TensorRT-LLM, GPUs (H100, A100), Cloud Run     |
+-------------------------------------------------------------------------------+

Key Concept: What is an AI Engineer?

Chip Huyen defines the AI Engineer as the bridge between raw foundation model capabilities and practical enterprise utility.

Traditional Software Engineers write deterministic logic (if-else). Traditional ML Engineers curate labeled features and train PyTorch architectures. AI Engineers compose pre-trained foundation models into resilient, multi-step systems using prompts, retrievers, tools, and evaluations.

The Four Stages of AI Adoption in Organizations

1. Stage 1: Consumer Exploration: Individual employees use web interfaces (ChatGPT, Claude.ai, Perplexity) for ad-hoc productivity.
2. Stage 2: Wrapper Applications: Simple API calls wrapping foundation models with system prompts and lightweight UI.
3. Stage 3: Enterprise Integration (RAG & Tooling): Connecting models to internal data lakes, CRM systems, databases, and permission models.
4. Stage 4: Custom Flywheels: Fine-tuning specialized internal models on proprietary transaction logs, user interactions, and curated domain data.

The "Jagged Frontier" of AI Capabilities

A crucial insight from Chapter 1 is that foundation model capability is not smooth:

• An LLM can write complex quantum physics simulations in Python, but struggle to count the number of 'r's in "strawberry".
• Systems succeed when engineers build deterministic scaffolding (code interpreters, regex parsers, schema validators) around the model's jagged blind spots.