20 LESSONS · ONE PRACTICAL PROJECT
Build your way into
GenAI engineering.
From your first policy file to a grounded assistant, a safe tool call, and a production plan. Follow a clear reading sequence, test the ideas, and keep the work you produce.
How to use this series
Read in order on your first pass. Each lesson has a worked explanation, a visual guide, a practice task, and a checkpoint. You will build around Northstar Supplies, a fictional support assistant with three policies and a read-only order lookup.
Bring basic computer skills and Python 3 for the coding exercises. Early lessons and the downloadable lab run without an API key. Lesson 8 offers an optional hosted API exercise that requires your own account and may incur charges. Read times cover the article; practice time depends on your experience.
The downloadable lab is a runnable lexical retrieval baseline with explicit limits. Later lessons explain how to add embeddings, generation, and deployment controls. It is a starting point for practice, not a ready-made production service.
01 / Foundations
Build the vocabulary and a working Python base.
Your GenAI engineering roadmap: build something worth using
Separate model capabilities from application responsibilities.
Read and practice →LESSON 02Generative AI fundamentals: what the model can and cannot know
Explain generation, inference, training, and grounding.
Read and practice →LESSON 03Python for GenAI: turn policy files into clean records
Use lists, dictionaries, functions, and exceptions in an ingestion task.
Read and practice →LESSON 04Machine learning and deep learning: learn from the right examples
Distinguish supervised, unsupervised, and reinforcement learning.
Read and practice →LESSON 05Transformers explained through one ambiguous sentence
Trace tokens through a transformer block.
Read and practice →LESSON 06Large language models: choose by behavior, not a leaderboard
Distinguish pretraining, adaptation, and inference.
Read and practice →02 / Retrieval and quality
Connect models to evidence, then measure the result.
Prompt engineering: write a contract the answer can satisfy
Write instructions with a clear task, evidence boundary, and output format.
Read and practice →LESSON 08LLM APIs and structured output: make the response usable
Call a model from a server-side Python script.
Read and practice →LESSON 09Embeddings: measure relatedness without mistaking it for truth
Explain what a text embedding represents.
Read and practice →LESSON 10Vector databases: store meaning with identity and permissions
Choose exact or approximate search for a measured need.
Read and practice →LESSON 11Build a RAG pipeline that can say “the policy does not cover that”
Separate offline ingestion from online answering.
Read and practice →LESSON 12Advanced RAG: repair retrieval one failure at a time
Combine lexical and semantic candidates.
Read and practice →LESSON 13Evaluate LLM and RAG systems with cases you can defend
Build a labeled set with ordinary and adversarial cases.
Read and practice →03 / Tools and responsible behavior
Add actions, memory, safeguards, and adaptation.
AI agents: give the loop a job, a budget, and an exit
Distinguish a fixed workflow from a model-directed loop.
Read and practice →LESSON 15Tool calling and MCP: connect a model without surrendering control
Describe the full tool-call round trip.
Read and practice →LESSON 16Memory and context: remember the useful parts, retire the rest
Separate request context, conversation history, and durable memory.
Read and practice →LESSON 17Guardrails and security: protect the boundary around the model
Identify untrusted inputs across retrieval and tools.
Read and practice →LESSON 18Fine-tuning: decide whether training solves the actual problem
Distinguish prompting, retrieval, and parameter adaptation.
Read and practice →04 / Delivery and demonstration
Plan a release and present a project you can defend.
What you will have at the end
A project brief, inspectable document records, a prompt and output contract, retrieval experiments, an evaluation set, an authorized lookup boundary, and a release plan. The final lesson helps you turn those artifacts into a reproducible demonstration.