A 3D map of the course: prompt, model, retriever, tools, and an eval gate. On a phone it waits until you open it. If motion is reduced, or this browser has no WebGL, the map stays flat.
Offline course · October 2026 · MIT code · CC BY-NC-SA content
Learn to build AI systems you can test.
A lifelong student of systems, networking, voice, and the contact center, studying AI omni-channel agents. The map is prompt, model, retrieval, tools, and an eval gate. No account. No key required to read it.
- Tracks
- 16
- Lessons
- 593
- Reading time
- 137 hours
- Network calls
- 0
Diagrams, flows, and charts
The same counts are on every track page.
Counted from lesson ids in this build. Rooms are the long imported pages.
Glossary, interview, and FAQ are reference sections and are not in these bars.
Minutes are the lesson field in the catalog, including expanded glossary stubs.
Levels are the course labels core, practice, and advanced.
Counted from check blocks on track lessons, not from the quiz bank.
Room diagrams are the 38 study-draft files. Animated flows are drawn for this campus.
The bar chart above is the same minute counts. This frame is filled by the vendored D3 script.
Rough.js draws this frame from the lesson count 593. The number is the catalog.
Architect conversation
Two short clips about how to view, practise, and share this lab. Two Piper voices, en_GB alba and en_US lessac, over a synthesized chord pad. The words are also in audio/scripts/.
Rate, comment, or help from the repository. Share the lab from the Labs link. This audio is a study guide, not a certificate.
Open the diagram atlas Play the flows Open the charts Cheat codes
Tracks
Take them in order the first time. Each card says why that track matters. The bar is your progress in this browser. The catalog count is every study page, including short glossary entries and interview drills. The 581 figure is 128 course lessons plus 453 imported pages, checked by id before the ecosystem track was added.
00Start hereLearn how this course works, what to save as proof, and the safety rules that apply everywhere else.Why it matters. Later tracks only work if you can keep proof and follow one safety rule.01How models workExplain tokens, context, sampling, hallucination, and the model families named in current job posts.Why it matters. Token, context, and sampling mistakes cause more bad product calls than a weak slogan.02PromptingWrite prompts as jobs with a done-check, an output shape, and a test for the next change.Why it matters. A prompt without a done-check cannot be improved, because nobody can say it failed.03Local modelsRun an open-weight model on your own machine, pick a size that fits, and keep data separate from weights.Why it matters. A local model is how you keep data off a hosted service and learn the real memory limit.04APIs and costCall a hosted model without leaking the key, and read a bill in tokens, cache, and retries.Why it matters. A leaked key or an unread bill ends a project faster than a clumsy prompt.05RetrievalBuild a retrieval pipeline that cites sources, and measure whether the right passage was found.Why it matters. An answer is grounded only when it cites a passage you actually retrieved.06AgentsDesign a tool loop with a schema, a permission boundary, a step cap, and a trace.Why it matters. A tool loop without a schema and a step cap spends money and takes actions you did not allow.07MCPExplain the Model Context Protocol and connect tools without granting a server silent access.Why it matters. A shared tool protocol helps only when the server cannot act without the permission you granted.08Coding agentsReview an agent that edits a repository: plan, diff, tests, and a human gate.Why it matters. A repository edit is finished when the diff and the tests agree with the task.09EvalsScore quality, safety, latency, and cost on a versioned task set before you trust a change.Why it matters. A change that feels better can still lose on quality, safety, latency, or cost.10Fine-tuningDecide when LoRA is the right next step, and compare an adapter to the base model on the same tasks.Why it matters. LoRA is warranted when the same eval says prompting and retrieval already failed.11SafetyWrite a personal policy for data, tools, and the actions that always need a person.Why it matters. The actions that need a person have to be written down before a tool can take them.12Product skillsSpecify an AI feature the way a product manager and an engineer can build and measure together.Why it matters. A feature ships when one spec is measurable by both the product owner and the engineer.13CapstoneFinish a small grounded assistant with citations, one tool, ten eval tasks, and a safety note.Why it matters. The course ends with a small assistant you can test, not a pile of notes.14Contact centerPlace a voice model beside the queue, the recording, and the human who still owns the call.Why it matters. A voice model that owns the queue will mishandle emergency calls and the recording.EXEcosystemMap sister systems, ports, cutovers, and the claims this course will not certify.Why it matters. Ports, cutovers, and sister systems are how a demo becomes a system you can operate.
No track is in that state yet.
Playgrounds
23 labs. Zero keys. The numbers are fixed teaching values. Each track has a hands-on one.
00Proof recordA run is not proof until it has a trace, a metric, and a file someone else can open.01Tokens and attentionSee which earlier token a position is allowed to use.01Temperature and top-pWatch one toy distribution get sharper, wider, and then cut by top-p.02Prompt diffSee which words turn a wish into a job, an input, and a done-check.03Memory fitTurn parameter count and bit width into a weight file, then see what still has to fit.04Cost and latencyPrice a call from tokens and a toy rate, then see what a cache does to the bill.04Context packingFill a 100-token window without dropping the question, the source, or room to answer.05Retrieval pipelineStep from chunks to vectors to a cited answer, or a refusal.05Embedding spaceDrag a query through a toy map and read the cosine to each note.05Vector indexCompare a flat scan with a grid that only opens nearby cells.06Agent and MCPWatch a tool call, then the gate that blocks a send.06Function-call checkAccept a tool call only when the JSON matches the schema.06Multi-agent handoffPass a source from a researcher to a writer, and stop when the source is empty.07MCP permissionsAllow a read, ask before a draft, and deny a send.08Diff reviewA repository edit is finished when the secret is gone and the test is named.09Eval gateMove quality, safety, latency, and cost until the gate says stop.10LoRA rankSee how rank changes the count of trainable numbers.10Prompt, retrieval, or fine-tunePick the cheapest fix the answers actually justify.11Guardrail sandboxSee a refusal rule fire. The lab never carries out the request.12One-page specName the user, the job, a real metric, and a non-goal.13Capstone checklistTests, a build log, a short demo, and a scorecard, or it is not ready.09Quiz and boss examCited questions across the tracks, plus multi-step incident exams.00FlashcardsThe check questions from the lessons, stored in this browser.
Continue
Nothing marked complete yet.
Marks
Complete a lesson to unlock the first mark. Marks stay in this browser.
Choose a route
Short paths through the same lessons. The track map is the full course.
New to AI
Enough mechanism to stop treating the model as a search box.
Grounded assistant
Retrieve, cite, call one tool, and prove the answers.
Coding agent work
The loop teams expect when a job post says coding agent or MCP.
Product and PM
Spec, metrics, failure states, and the build-or-buy choice.
Voice agent bridge
The model sits beside the phone system. It does not own the call.
How a tested answer is built
Play the path. Each step is a habit the tracks practice.
Animated flowFrom a prompt to a gate
Transcript. Welcome to AI Lab Free University. This is an offline course in models, retrieval, agents, evals, and the judgment to ship or stop. Progress stays in this browser.
Uses this device's speech engine when it has one. Nothing is uploaded.
Esc
↑↓ move Enter open Esc close
Audio
0:00