Building with LLMs
Use foundation models as a building block — prompting, retrieval-augmented generation, agents, and the architecture of a production LLM application.
14 lessons
01Prompt Engineering & In-Context Learning
How a frozen model adapts to your task from the prompt alone — zero/few-shot, the anatomy of a prompt, and the decoding knobs that shape every answer.
Chain-of-Thought & Structured Output
Two opposite needs: letting the model reason freely across intermediate steps, and forcing its final answer into a schema your code can parse.
Embeddings & Semantic Search
Turning text into vectors whose geometry encodes meaning — cosine similarity, vector indexes, and the retrieval engine that powers RAG.
Retrieval-Augmented Generation
Grounding an LLM in your own documents — chunk, embed, retrieve, rerank, and generate — plus where RAG fails and how to evaluate it.
Agents & Tool Use
Letting an LLM act — function calling, the ReAct loop of Thought/Action/Observation, memory, and why agents are powerful but hard to make reliable.
AI-Engineering Architecture
What wraps the model in production — guardrails, prompt-injection defense, model routing, caching, observability, and the user-feedback loop.
Reasoning Models
How o1, DeepSeek-R1, and Gemini Flash Thinking differ from base LLMs — chain-of-thought training via RL, test-time compute scaling, and when to use a reasoning model.
LLM Evaluation & LLM-as-a-Judge
Why you can't ship an LLM app on vibes: offline eval suites, reference-based vs reference-free metrics, LLM-as-a-judge with its biases, and the regression testing that catches a prompt change that quietly breaks production.
LLM Observability, Tracing & Prompt Management
Operating an LLM app you can actually debug: traces and spans across a multi-step call, the token/cost/latency metrics that matter, output drift, and treating prompts as versioned, tested artifacts.
Guardrails, Safety & LLMOps Security
The defensive layer of a production LLM app: input and output guardrails, the OWASP LLM Top 10 (prompt injection, data leakage, excessive agency), and the defence-in-depth posture that assumes every input is hostile.
Building with LLMs Quiz
Test your understanding of prompting, decoding, semantic search, RAG, agents, reasoning models, evaluation, and LLM application architecture.
Code Intelligence & Generation
How GitHub, Uber, and Meta build LLM-powered coding tools — code embeddings and search, fill-in-the-middle completion, repo-level context, code review automation, and test generation.
Voice & Multimodal AI
End-to-end voice AI pipelines, vision-language models, and multimodal RAG — the architectures behind Intercom's voice agent, Instacart's visual catalog, Airbnb's speech search, and Walmart's shopping assistant.
RAG Architectures: Which Kind of RAG Do You Need?
The nine families of retrieval-augmented generation — naive, hybrid, query-transforming, hierarchical, graph, self-correcting, agentic, multimodal, and structured — what each one fixes, what it costs, and how to pick.