AI Playground

Courses

Learn AI by building it

26 missions across 6 courses. Start with Foundations — half an hour and you’ll have built a working agent. The rest go deep on whatever you need.

Prefer to look something up? Browse the concept reference →

0 of 26

AI Foundations

0/6

Build a working AI agent from nothing, and understand every part.

Afterwards: Explain what an LLM, RAG, a tool, an agent and an eval actually are.

  1. 1Make AI TalkBuild the smallest possible AI, and run it.InputLLMOutput3–5 min
  2. 2Control Your AISame question, completely different answer.PromptingSystem instructions4–6 min
  3. 3Give AI KnowledgeCatch your AI making something up, then fix it.RAGRetrievalGroundingHallucination6–8 min
  4. 4Give AI ToolsSome jobs shouldn't go to a language model at all.ToolsFunction calling5–7 min
  5. 5Build an AgentStop wiring the path. Let it choose.AgentsGoalsTool selection6–8 min
  6. 6Make Your AI BetterFind the problem, measure it, fix it, measure again.ObservabilityEvalsIteration8–10 min

Prompting & Context

0/4

Getting the right things in front of the model at the right moment.

Afterwards: Shape output with examples, reasoning and structure instead of hoping.

  1. 1Show, Don't TellThree examples beat three paragraphs of description.Few-shot promptingIn-context learning5–7 min
  2. 2Make It Show Its WorkingAnswering immediately is how models get multi-step problems wrong.Chain of thoughtReasoning models5–7 min
  3. 3Answers Code Can UseProse is lovely for humans and useless for software.Structured outputSchemasExtraction5–7 min
  4. 4What It Can Hold in MindTokens, context windows, and why long prompts get expensive.TokensContext windowContext engineering6–8 min

Knowledge & Retrieval

0/4

Take RAG apart and tune every stage of it.

Afterwards: Diagnose why a retrieval system is returning the wrong passages.

  1. 1Take RAG ApartOne Knowledge block is really six stages in a coat.ChunkingEmbeddingsVector storesRetrieval8–10 min
  2. 2Chunking Changes EverythingThe least glamorous setting in RAG, and one of the most important.Chunk sizeOverlapRetrieval quality6–8 min
  3. 3Search Two WaysMeaning search misses exact strings. Keyword search misses synonyms.Hybrid searchBM25Reranking7–9 min
  4. 4Show Your SourcesAn answer nobody can check is an answer nobody should trust.CitationsFaithfulnessHallucination detection7–9 min

Agents & Tools

0/4

Systems that plan, delegate, connect to real services, and know when to ask.

Afterwards: Design an agent that's capable without being dangerous.

  1. 1Make a Plan FirstImprovising a six-step job rarely ends well.PlanningTask decomposition6–8 min
  2. 2Plug In a Real ServiceOne standard plug instead of an integration per model.MCPTool schemasInteroperability6–8 min
  3. 3Ask Before You ActCapability is easy. Knowing when to stop is the hard part.Human in the loopAutonomyBlast radius6–8 min
  4. 4A Team of AgentsSometimes. Often one good agent is better than three confused ones.Multi-agentDelegationReflection8–10 min

Quality & Safety

0/4

Measuring whether it works, and finding out how it breaks.

Afterwards: Build a test set, grade with a judge, and attack your own system.

  1. 1A Real Test SetTrying it once and feeling good is not testing.Golden datasetsRegression testing6–8 min
  2. 2Let AI Grade AIPowerful, scalable, and quietly biased.LLM-as-judgeRubricsJudge bias6–8 min
  3. 3Hide the Personal DetailsEverything you send a model may end up in a log.PIIRedactionData privacy5–7 min
  4. 4Break Your Own AISomeone is going to try. Much better that it's you.Prompt injectionRed teamingTrust boundaries8–10 min

Production

0/4

Surviving real traffic, real budgets and real failures.

Afterwards: Cut cost and latency without quietly cutting quality.

  1. 1Don't Pay TwiceReal traffic is astonishingly repetitive.CachingSemantic cacheLatency6–8 min
  2. 2When Things BreakModel APIs go down. Your product doesn't have to.RetriesFallbacksGraceful degradation6–8 min
  3. 3The Right Brain for the JobMost requests are easy. Stop paying frontier prices for them.Model routingCost optimisationClassification6–8 min
  4. 4Close the LoopYour users are running the best eval you'll ever have.User feedbackVersioningThe data flywheel7–9 min

Progress is saved on this device. Accounts come later.