Skip to content
Learn Gen AI engineering

18 parts · 266 topics · 1,916 chapters

Learn Gen AI engineering

Each topic is a working engineer's account of one problem: what breaks, why, and what to do about it. Chapters are ordered so each follows from the last.

Mid 89 topics · Senior 88 topics · Lead 89 topics

This week

0 / 10

topics read

Your first chapter starts the week.

Start at the beginningPython dependency lockfiles

Python Engineering0 of 2 topics

Part 1 · 3 topics · 21 chapters

Working Foundations

Deliberately tiny. This is the tooling and the arithmetic, not a maths course.

Python Engineering

Build and organize Python programs for AI work.

Vector Arithmetic

Compute and compare vectors using basic operations.

Part 2 · 18 topics · 136 chapters

LLM Mechanics at Inference

How the thing behaves when you call it. Operational mechanics, not architecture theory.

Token Context

Understand how token representation and context limits shape inference behavior.

Generation Mechanics

Control generation and reason about its randomness, stopping behavior, latency, and caching.

Knowledge Boundaries

Distinguish supplied information, learned knowledge, and fabricated model output.

Model Variants

Choose among model types and understand their trade-offs and specialized inference modes.

Part 3 · 22 topics · 159 chapters

Model APIs

The daily job, and the closest thing to your existing skill set.

Request Construction

Build well-formed, multimodal, and controllable model requests.

Conversation State

Manage conversation history, persistence, and provider-hosted state.

Request Delivery

Integrate, stream, operate, scale, and test model API calls.

Cost Control

Budget tokens, use prompt caching, and attribute model spend.

Provider Strategy

Coordinate providers, platform constraints, fallbacks, and model lifecycle.

Part 4 · 17 topics · 121 chapters

Evaluation

Before prompting, on purpose. This is what separates mid from senior, and everything downstream is guesswork without it.

Evaluation Basics

Understand why generative systems need probabilistic evaluation rather than ordinary exact-match testing.

Dataset Diagnosis

Build representative evaluation data and use failures to understand and prioritize system weaknesses.

Scoring Methods

Apply deterministic checks, similarity metrics, and carefully calibrated model judges.

Eval Operations

Select evaluation tooling and integrate cost-aware quality gates into development workflows.

Part 5 · 19 topics · 141 chapters

Prompting & Context Engineering

Prompts are versioned, tested artifacts with an interface contract. Context engineering is the senior half of this stage.

Prompt Design

Construct prompts with clear structure, examples, placement, decomposition, and model-aware adaptations.

Reasoning Methods

Use scratchpads, self-critique, and sampling strategies to improve reasoning reliability.

Prompt Lifecycle

Safely template, version, review, and automatically optimize prompts as production artifacts.

Context Control

Budget, order, select, compress, and measure context while managing conflicting instructions.

Part 6 · 15 topics · 109 chapters

Structured Output & Tool Calling

Turning the model into a typed component. Where your backend instincts get their first real foothold.

Typed Outputs

Produce, validate, repair, stream, and verify structured model responses.

Tool Invocation

Execute model-selected tools through reliable calls, results, and streaming arguments.

Tool Design

Design tool surfaces, descriptions, results, semantics, versioning, and failure handling.

Part 7 · 16 topics · 114 chapters

Model Context Protocol (MCP)

The standard interface between agent applications and the tools, data and prompts they use. Its own stage because it is now the default integration surface, and because most of what goes wrong with it is auth and trust, not protocol.

Core Protocol

Understand MCP’s participants, lifecycle, and connection transports.

Protocol Surfaces

Use MCP capabilities to expose tools, context, prompts, and interaction flows.

MCP Integration

Build, consume, debug, and assess MCP integrations in real systems.

Trust Controls

Secure MCP access and govern the risks of multi-tenant and third-party servers.

Part 8 · 11 topics · 73 chapters

Prompt Injection & Tool Safety

Kept deliberately (see Scope). Not a security specialism — the architectural constraints that determine how an agent can be built at all.

Injection Threats

Recognize how untrusted content, model outputs, and tool interfaces create injection and exfiltration risks.

Safe Execution

Design constrained, observable, and resource-bounded agent execution with appropriate human oversight.

Part 9 · 19 topics · 135 chapters

Retrieval

Measured before it is optimized. Most bad RAG is bad search, not bad prompting. Lexical first, deliberately.

Search Foundations

Define relevance, build a lexical baseline, and measure retrieval quality.

Embedding Models

Understand, select, and configure embeddings for semantic and multilingual search.

Vector Infrastructure

Build, filter, combine, compress, and maintain production vector indexes.

Result Quality

Improve retrieved results through reranking, richer matching, deduplication, and domain adaptation.

Part 10 · 21 topics · 163 chapters

RAG

Retrieval with a generator bolted on, and every failure is still a retrieval failure. Build the naive version first so you have something to measure, then fix what the measurements actually say — not what the blog posts say.

Pipeline Foundations

Build and operate the core document-to-answer pipeline.

Retrieval Design

Shape queries and choose retrieval architectures for different information needs.

Evaluation Tuning

Measure RAG quality and systematically improve its parameters.

Grounded Answers

Ground answers, manage uncertainty, and prevent unreliable or unauthorized results.

Multimodal Retrieval

Parse and retrieve information from images, charts, tables, and visual documents.

Part 11 · 20 topics · 148 chapters

Agents: The Core Loop

One agent, one loop, done properly. Architectures come next, and only make sense once this is solid.

Core Loop

Define, implement, and reason about the basic agent control loop.

Reliable Execution

Make agent execution durable, bounded, concurrent, and practical to operate.

Oversight & Evaluation

Control agent authority and measure behavior, quality, cost, and failure recovery.

Memory Policies

Design, protect, and evaluate what agents remember across sessions.

Part 12 · 18 topics · 132 chapters

Agentic Architectures & Multi-Agent Patterns

The named topologies, what each is actually for, and the failure mode each introduces. The recurring senior judgment is that **most of these are more expensive than the single loop and only some of them are better.**

Workflow Design

Choose and compose simple fixed workflows for the task shape.

Reasoning Patterns

Improve outputs through critique, verification, and adversarial comparison.

Agent Topologies

Structure planning, delegation, handoffs, shared artifacts, and isolated sub-agents.

Coordination Control

Manage communication, termination, cost, failure, and architecture-level evaluation.

Part 13 · 13 topics · 94 chapters

Agent Extensibility: Skills, Config & Subagents

How an agent acquires new capability and standing context without being rebuilt. This is where a lot of current practical leverage lives, and it is almost entirely undocumented as a discipline.

Skill Design

Choose, author, and validate reusable capabilities that agents can load when needed.

Standing Context

Configure durable, scoped instructions that give agents reliable project context.

Agent Operations

Delegate bounded work, enforce invariants, and distribute extensibility assets as maintained code.

Part 14 · 11 topics · 80 chapters

Coding Agents

Its own stage because it is the largest commercial application category, it has the cleanest verifiers of any domain, and it is where a backend engineer's existing skills convert at the highest rate.

Codebase Grounding

Ground the agent in repository structure, conventions, context, and deterministic verification.

Verified Changes

Create, isolate, test, evaluate, and manage code changes through feedback-driven loops.

Part 15 · 6 topics · 41 chapters

Data & Annotation for AI

Only the parts that are AI-specific. Your existing data engineering — cleaning, lineage, orchestration, CDC, data contracts — transfers unchanged and is not re-taught.

Evaluation Datasets

Build reliable evaluation data through annotation, synthetic generation, and reproducible versioning.

AI Data Pipelines

Prepare production corpora efficiently while filtering low-quality content and protecting personal data.

Part 16 · 18 topics · 118 chapters

Production LLMOps

Your existing craft applied to a new failure profile. Restricted to the *delta* from ordinary SRE — capacity planning, canaries and runbooks are assumed.

Platform Foundation

Design and operate the gateways, providers, stores, quotas, and capacity behind LLM traffic.

Runtime Assurance

Instrument production behavior, protect operational data, and respond to AI-specific incidents.

Efficiency Control

Reduce LLM cost and latency through caching, routing, batching, and response optimization.

Quality Lifecycle

Safely release changes, evaluate outcomes, learn from users, and detect quality drift.

Part 17 · 7 topics · 49 chapters

Multimodal & Voice

Image *input* is in Stage 2 and document parsing in Stage 9, because RAG needs both. This is the rest of the modality surface.

Image Systems

Build systems that understand, retrieve, secure, generate, and edit visual content.

Voice Systems

Build speech pipelines and realtime agents for natural voice interaction.

Part 18 · 12 topics · 82 chapters

Senior Scope

None of these are about writing more code.

Opportunity Strategy

Choose worthwhile AI problems, product surfaces, prototypes, and economic models.

Trustworthy Products

Design AI experiences that make uncertainty, errors, and user control explicit.

AI Stewardship

Sustain technical judgment, portability, research awareness, and organizational standards.

Back to top