Go beyond "what is AI" into how it actually works under the hood, the real toolbox of AI products worth knowing, and the vocabulary that lets you follow any AI conversation with confidence.
🧭 Builds on Beginner
🖱️ Click to Interact
⏱️ ~100 min session
Use → / ← arrow keys, scroll, or the buttons below to navigate. Look for dashed buttons and clickable cards throughout — this deck talks back.
🧠
🔧
🔗
💬
🧩
📊
🤖
🖱️ drag to rotate — tools, mechanics & terms orbiting the model
Welcome
What You'll Walk Away With
Click a card to flip it and see a fun fact.
🧰
A real toolbox, not just chatbotsCoding agents, automation platforms, research tools, and how they connect.
FUN FACT
The AI coding-tools market alone grew from $4.9B in 2024 to a projected $9.46B in 2026 — this space is moving fast.
⚙️
How the model actually worksTokens, embeddings, attention, context windows — demystified.
FUN FACT
A "token" usually isn't a whole word — long or rare words often get split into two or three pieces.
📖
The vocabulary that unlocks everything elseRAG, fine-tuning, agents, MCP — said with confidence, not guesswork.
FUN FACT
Anthropic's Model Context Protocol (MCP) now has over 10,000 active public servers and is stewarded by the Linux Foundation.
🗺️
Your path to AdvancedWhat "hands-on building" looks like once concepts click.
FUN FACT
Most people who understand these concepts well start prompting noticeably better within the same week.
The Big Picture
Your Learning Roadmap
Today we focus on Intermediate — with a quick recap of Beginner and a preview of Advanced. Click a level to see what's inside.
Level 1
🌱 Beginner
What is AI, how it works, and your first AI tools.
AI basics · Types of AI · 6+ tools compared · Your first prompts
→
Level 2 · Today
⚙️ Intermediate
The AI toolbox, how models work, and essential terminology.
Tool categories · Tokens, embeddings & context · RAG vs. fine-tuning · The glossary everyone should know
→
Level 3
🏗️ Advanced
APIs, automation platforms, and connecting AI to real business data.
Building with APIs · Deploying agents · Real RAG pipelines · Fine-tuning basics
→
Level 4
🏆 Expert
Fine-tuning, production AI systems, and governance at scale.
MLOps · Evaluation & governance · Leading AI strategy
Module 1 · The AI Tool Landscape
The Six Categories of AI Tools
Click a category to see why it belongs on your radar.
Module 1 · Building Your Own Assistant
Custom GPTs, Claude Skills & Gemini Gems
Every major chat assistant now lets you build a specialized, reusable version of itself — no code required. Click a tab to see how each one does it.
Package a persona + instructions + files
Give it a name, a system prompt, and reference documents — it becomes a shareable, purpose-built assistant other people can use too.
Reusable skills and persistent project context
A Skill packages instructions and know-how for a repeatable task; a Project keeps files and context loaded across every conversation inside it.
A saved persona with its own instructions
Define a Gem once — tone, expertise, constraints — and reuse it any time instead of re-explaining context in every new chat.
The pattern is the same everywhere: stop re-typing the same context every conversation — save it once as a reusable assistant instead.
Module 1 · Building With AI
💻 AI Coding Assistants
Three different philosophies for the same job — click a tab.
An AI-native code editor
Fast autocomplete and in-editor chat built for flow — great for quick, low-friction edits to small and medium tasks.
cursor.com
A terminal-first autonomous agent
Wins on deep reasoning — debugging, multi-file refactors, and architectural changes it can plan and execute on its own.
claude.com/claude-code
Deep GitHub integration
The broadest IDE support and the tightest fit with pull requests, code review, and CI — good for teams already living in GitHub.
github.com/features/copilot
The most common 2026 developer setup is a two-tool stack: an in-editor assistant for daily work, plus a terminal-based agent for deep multi-file tasks.
Module 1 · A Key Distinction
What Makes a Tool "Agentic"?
Agentic means the AI doesn't just respond — it plans, takes actions, checks results, and adjusts, often across several steps without you approving each one.
💬
A chatbot answersYou ask, it replies. One turn, no independent action.
▾
Classic ChatGPT/Claude/Gemini chat: you're the one driving every step.
🤖
An agent actsGiven a goal, it decides the steps and executes them itself.
▾
"Book the cheapest flight under $400" — an agent searches, compares, and books, checking in only when it needs a decision from you.
🛠️
Tool use is what makes it possibleAgents call real tools — search, code execution, APIs — not just text generation.
▾
This is where "function calling" and MCP (more on that later) come in — a standard way for a model to say "run this tool" and get a real result back.
⚠️
More autonomy, more responsibilityThe more a tool can do on its own, the more it needs oversight.
▾
We'll come back to this in Module 4 — permissions, reviewing agent actions, and staying in the loop on anything high-stakes.
🧭Plan
🔎Search
🛠️Act
✅Verify
🎉 That's an agent loop — click each step in order.
Module 1 · Beyond the Chat Window
Browser & Computer-Use Agents
Tap a card to see what it actually does.
🌐Browser agents
AI that can navigate real websites — click, fill forms, and read pages — on your behalf.
Claude in Chrome and similar tools can research a topic across multiple tabs and summarize findings without you doing the clicking.
👆 tap for example
🖥️Computer-use agents
AI that can operate a whole desktop — open apps, move the mouse, type — not just a browser tab.
Useful for repetitive cross-app tasks: pulling a number from a spreadsheet and pasting it into a native app your company uses.
👆 tap for example
🔐Permission tiers
Good agent tools distinguish "can look" from "can click" from "can type" — not all-or-nothing access.
This matters for safety: a tool that can read your screen shouldn't automatically be able to submit forms or send messages.
👆 tap for example
👀Why humans still watch
These agents are powerful but not perfect — a wrong click on the wrong page can matter.
Best practice: let agents handle research and repetitive steps, but review before anything gets submitted, purchased, or sent.
👆 tap for example
Module 1 · Connecting the Dots
Automation Platforms + AI
Step through how a no-code automation actually runs.
⚡
1. Trigger
Something happens that starts the workflow. e.g. a new form response comes in, or a new email arrives.
🧠
2. AI step
An AI model processes the data — drafting, classifying, summarizing, or deciding what happens next. e.g. it drafts a reply in your tone, or tags the request by urgency.
✅
3. Action
The platform carries out the result automatically in another app. e.g. the drafted reply lands in your inbox, or the ticket gets routed to the right team.
Platforms like Zapier, Make, and n8n popularized this pattern — you rarely need to write code to wire an AI step into a real workflow.
Module 1 · At Work, Beyond Chatbots
📈 Enterprise AI Platforms
Click a card to flip it and see what it's good for.
🔎
GleanUnified AI search across Slack, Drive, Jira, Salesforce, and 90+ other connected tools.
GOOD FOR
Asking one question and getting a ranked answer pulled from every tool your company uses — respecting existing permissions. glean.com
☁️
Salesforce AgentforceAgentic AI built directly into CRM workflows.
GOOD FOR
Sales and service teams who live inside Salesforce daily — deep CRM depth, though largely confined to that platform. salesforce.com/agentforce
💬
Slack AISummarizes threads and channels, surfaces what you missed.
GOOD FOR
Catching up on a busy channel in seconds instead of scrolling. slack.com
Notice the shift from Module 1's productivity tools: these platforms don't just draft text — they act on and across your company's real systems.
Module 1 · Research & Knowledge
🔎 Research & Knowledge Tools
Flip the switch to compare two different jobs.
Perplexity — research assistant
Every answer links back to real, current sources — built for fact-checking and competitive research.
perplexity.ai
NotebookLM — grounded in your files
Upload your own PDFs, notes, or slides and the AI answers only from that material — with citations back to your source.
notebooklm.google.com
Both solve the same underlying problem — keeping AI answers grounded in real sources instead of just "what the model remembers."
Module 1 · Knowledge Check
Quick Check: Match the Tool to the Job
1. You need an AI that can plan and execute several steps on its own, like booking a flight. What's this called?
A chatbot
An agent
A vector database
2. You want AI answers grounded only in your own uploaded documents. Which type of tool fits best?
A "chat with your documents" tool like NotebookLM
A general chatbot with no uploads
An image generator
Module 1 · Zooming Out
Evolution of AI Tools: A Timeline
Click a milestone for more detail.
🖱️ drag to spin the dial
2017 — the transformer architecture is published.
Module 2 · How AI Actually Works
From Words to Numbers: Tokens
Models don't read letters or words the way you do — they read tokens, small chunks of text converted into numbers. Try it below.
Simplified educational demo — real tokenizers split words into smaller sub-word pieces, so actual token counts usually run higher than this.
Module 2 · How AI Actually Works
Embeddings: Turning Meaning Into Math
Each token gets converted into a long list of numbers — an embedding — that represents its meaning as a position in space. Words used in similar ways end up near each other.
📍
Meaning becomes a location"King" and "queen" land closer together than "king" and "banana."
▾
This is why models can reason about analogies and synonyms — geometrically close embeddings behave like related concepts.
🗂️
This is also how search improvedVector search finds results by meaning, not just matching keywords.
▾
A search for "cheap flights" can now match a document that says "budget airfare" — no shared keyword needed.
👑 royalty
🍎 fruit
king
queen
banana
apple
🖐️ drag each word to where it belongs
Module 2 · How AI Actually Works
Inside the Model: Attention & Transformers
The transformer architecture — behind GPT, Claude, and Gemini alike — uses a trick called attention to decide which earlier words matter most for predicting the next one.
👁️
Attention weighs every earlier wordNot just the word right before — the whole sentence, weighted by relevance.
▾
In "the trophy didn't fit in the suitcase because it was too big," attention helps the model figure out "it" means the trophy, not the suitcase.
⚡
It runs in parallel, not word-by-wordThis is what made transformers so much faster to train than older architectures.
▾
Older models (RNNs) processed text one word at a time in sequence. Transformers process a whole chunk at once — a huge speed unlock.
Nobody hand-coded grammar rules into these models — the patterns emerged from predicting text at massive scale.
Module 2 · How AI Actually Works
The Context Window
The context window is how much text a model can "hold in mind" at once — your whole conversation, any documents you've shared, and its own reply, all counted together. Try filling one up:
0 / 8 slots used — click "Add a message" above
Module 2 · How AI Actually Works
Training vs. Inference
Flip the switch to compare the two phases.
Happens once, before you ever chat
The model studies enormous amounts of text and adjusts billions of internal values until it gets good at predicting the next token. Expensive, slow, done by the AI company.
Happens every time you send a message
The already-trained model uses what it learned to generate a response to your specific prompt. Fast, cheap per-request — this is what you experience as "using AI."
When people say a model was "trained on data up to a certain date," that's the training phase — inference is everything that happens after, including in this very conversation.
Module 2 · How AI Actually Works
Temperature & Sampling: Why Answers Vary
At each step the model doesn't pick one "correct" next token — it picks from a probability distribution. Temperature controls how safe or adventurous that pick is.
🧊
Low temperaturePicks the most likely token almost every time — consistent, predictable.
▾
Good for factual Q&A, code, and anything where you want the same reliable answer twice.
🔥
High temperatureGives less-likely tokens a real chance — more varied, more creative.
▾
Good for brainstorming and creative writing — riskier for facts, since it's more willing to wander.
0.20
🖱️ drag the dial
"The capital of France is Paris."
Module 2 · How AI Actually Works
Why AI Hallucinates
Tap a card to see the honest explanation.
🎲It's always predicting, never "looking up"
Unless it's using a tool like search, the model is generating the statistically likely next token — not retrieving a stored fact.
This is why it can state a wrong date or citation just as fluently and confidently as a correct one — fluency isn't the same as accuracy.
👆 tap for the honest version
🕳️Gaps get filled with plausible guesses
When it doesn't truly "know" something, it tends to generate something that sounds right rather than saying "I'm not sure."
Newer models are explicitly trained to hedge more and cite sources — but the underlying tendency never fully disappears.
👆 tap for the honest version
🔗Grounding tools reduce it, don't eliminate it
Web search, RAG, and citations all help — because the model has real text to lean on instead of guessing.
This is exactly why "cited answers" tools like Perplexity exist — verifiable sources are the fix, not a smarter guess.
👆 tap for the honest version
✅Your job: verify what matters
Treat confident-sounding AI output on facts, numbers, and citations the way you'd treat a smart but occasionally wrong colleague.
Higher stakes = more verification. A blog draft can tolerate more risk than a legal clause or a medical claim.
👆 tap for the honest version
Module 2 · How AI Actually Works
RAG: Retrieval-Augmented Generation
Step through how a RAG answer actually gets built.
❓
1. Query
You ask a question — the system doesn't answer yet. e.g. "What's our refund policy for enterprise customers?"
🔎
2. Retrieve
Your documents were pre-split into chunks and converted to embeddings ahead of time. Now it embeds your question too, and finds the closest-matching chunks. e.g. it pulls the exact paragraph from your policy doc — not by keyword match, but by meaning.
➕
3. Augment
Those retrieved chunks get inserted into the prompt as extra context, right alongside your question. The model now "sees" the real policy text, not just what it memorized during training.
✍️
4. Generate
The model writes an answer grounded in that retrieved text — often with a citation back to it. This is why RAG-backed tools can answer accurately about your own private documents.
📄 Doc
🧠
🖐️ drag the document into the model
Module 2 · A Common Mix-Up
RAG vs. Agentic AI
These get lumped together constantly, but they answer two completely different questions. RAG asks "does the model have the right facts?" Agentic AI asks "can the model take multi-step actions on its own?" A system can be either, neither, or both.
Tap a card to see a concrete example.
🔎RAG alone
Grounds one answer in real documents. Single retrieval, single reply — no independent multi-step action.
A support bot that answers "what's our refund policy?" by retrieving the actual policy paragraph — one lookup, one grounded answer.
👆 tap for example
🤖Agent alone
Plans and acts across several steps, but isn't necessarily grounded in your private documents — it may lean only on what it already knows.
An agent booking a flight: it searches, compares options, and books — multiple autonomous steps, but no document retrieval involved.
👆 tap for example
🔎🤖RAG + Agent, together
Most serious real-world assistants combine both: an agent that can call RAG retrieval as just one tool, on any step of a longer plan.
A research agent that plans a report outline, retrieves cited facts via RAG for each section as it goes, then assembles the final draft — grounding and autonomy working together.
👆 tap for example
Rule of thumb: if someone asks "is this a RAG system or an agent," the honest answer is often "both — RAG grounds what it knows, agentic behavior is what it does with that knowledge."
Module 2 · Same Start, Different Shape
Two Flows, Both Starting From a Prompt
Watch the arrows — Agentic AI loops, RAG runs straight through once.
Module 2 · Choosing Your Approach
Fine-Tuning vs. Prompting vs. RAG
Three ways to make AI better at your specific job — click a tab.
Fastest, cheapest, no setup
Just write better instructions in the prompt itself. Good first move for almost everything — try this before anything more complex.
When the model needs facts it doesn't have
Connect it to your documents or database so answers are grounded in real, current, private information — no retraining required.
When the model needs a new skill or style
Retrain it on examples of exactly the behavior you want — consistent tone, a specialized format, a niche task. Slower and more expensive; usually a last resort.
Rule of thumb: try prompting first, add RAG when facts are the problem, and reach for fine-tuning only when neither solves it.
Module 2 · How AI Actually Works
Multimodal AI: Beyond Text
Click a card to see what it means in practice.
🖼️Vision
The model can read images — screenshots, photos, charts, handwriting — not just generate them.
Upload a screenshot of an error message and ask "what's wrong here?" — no typing the error out required.
👆 tap for example
🔊Audio
Some models can listen and speak directly, without a separate speech-to-text step.
Real-time voice assistants that hear tone and pauses, not just transcribed words.
👆 tap for example
🎬Video
Understanding or generating moving images with consistent objects and motion over time.
Ask a model to summarize the key moments of a video file you upload.
👆 tap for example
💻Code & structured data
Reading and writing code, spreadsheets, and structured formats as fluently as prose.
This is what lets a single model both write your report and the spreadsheet formula behind it.
👆 tap for example
Module 2 · Knowledge Check
Quick Check: How AI Works
1. Which phase happens every single time you send a message?
Training
Fine-tuning
Inference
2. Your AI needs to answer accurately from your private company documents. What's the best first approach?
Raise the temperature
Set up RAG so it retrieves from your documents
Just ask it more nicely
Module 3 · Essential AI Terminology
Glossary: Core LLM Terms
Click a term to expand it.
1🔤
LLM (Large Language Model)The general category ChatGPT, Claude, and Gemini all belong to.
▾
A neural network trained on huge amounts of text to predict and generate language — "large" refers to the number of parameters, often in the billions.
2🎛️
ParameterAn internal adjustable value the model learned during training.
▾
A model "size" like 70B means roughly 70 billion of these values — more isn't automatically better, but it's a rough proxy for capacity.
3⚡
InferenceThe act of running the trained model to generate a response.
▾
Every reply you get is one inference pass — this is also the phase that costs money per request via an API.
4🔌
APIThe programmable "back door" that lets software talk to a model directly.
▾
Instead of typing into a chat window, a developer sends a request in code and gets a response back — this is how AI gets built into other products.
Module 3 · Essential AI Terminology
Glossary: Prompting Terms
Click a term to expand it.
1🎯
System promptHidden instructions set before your conversation starts.
▾
"Custom instructions" and "Projects" features are really just an easy way to set a persistent system prompt.
20️⃣
Zero-shot / Few-shotAsking with no examples vs. giving a couple of examples first.
▾
Few-shot prompting ("here are 2 examples of the format I want, now do a third") often boosts consistency dramatically.
3🧵
Chain-of-thoughtAsking the model to reason step-by-step before answering.
▾
"Think through this step by step" often improves accuracy on math and logic — some models now do this automatically as "extended thinking" or "reasoning" modes.
4🌡️
TemperatureA setting that controls how predictable vs. varied the output is.
▾
Covered in depth back in Module 2 — low for facts and code, higher for brainstorming.
Module 3 · Essential AI Terminology
Glossary: Model Behavior Terms
Click a term to expand it.
1🌀
HallucinationConfidently stating something false.
▾
Covered in depth in Module 2 — the model is predicting plausible text, not looking facts up, unless it's using a grounding tool.
2🧭
AlignmentHow well a model's behavior matches what humans actually want.
▾
A "well-aligned" model is helpful, honest, and avoids harmful outputs — this is an active, ongoing area of AI research, not a solved problem.
3🚧
GuardrailsRules and filters that keep a model's outputs within safe, intended limits.
▾
Can be built into the model itself, or added on top by the company deploying it (e.g. blocking certain topics for a customer-support bot).
4⚖️
BiasSkewed patterns inherited from training data.
▾
If training text over-represents certain viewpoints or groups, the model's outputs can reflect that imbalance — part of why human review still matters.
Module 3 · Essential AI Terminology
Glossary: Data & Retrieval Terms
Click a term to expand it.
1📍
EmbeddingText converted into a list of numbers representing its meaning.
▾
Covered in Module 2 — the foundation that makes semantic search and RAG possible.
2🗄️
Vector databaseA database built to store and search embeddings by similarity.
▾
This is the piece that makes RAG's "retrieve" step fast, even across millions of documents.
3🔎
RAGRetrieval-Augmented Generation — grounding answers in real, retrieved text.
▾
Covered step-by-step in Module 2 — query, retrieve, augment, generate.
4🪟
Context windowHow much text the model can hold in one conversation.
▾
Measured in tokens — once it's full, the oldest content typically gets dropped to make room for new messages.
Module 3 · Essential AI Terminology
Glossary: Agent & Automation Terms
Click a term to expand it.
1🤖
AgentAn AI that plans and takes multi-step actions toward a goal.
▾
Covered in Module 1 — the key difference from a plain chatbot is independent action, not just conversation.
2🛠️
Tool use / function callingA model's ability to call a real tool and use its result.
▾
Instead of guessing an answer, the model can call a calculator, a search engine, or your company's API and use the real result.
3🔗
MCP (Model Context Protocol)A shared standard for connecting AI to tools and data.
▾
Think of it as a universal adapter — before MCP, every AI product needed custom-built code for every tool it connected to. One MCP server for, say, Notion gets built once, and any MCP-aware agent can use it — no bespoke glue code per pairing. Introduced by Anthropic in 2024 and now stewarded by the Linux Foundation with OpenAI, Google, AWS, and Microsoft all on board — over 10,000 active public MCP servers exist as of 2026.
4🕸️
OrchestrationCoordinating multiple AI steps or agents into one workflow.
▾
The automation-platform pattern from Module 1 — trigger, AI step, action — is a simple form of orchestration.
Module 3 · Essential AI Terminology
Glossary: Access & Model Types
Click a term to expand it.
1🔓
Open-source vs. closed modelCan you download and run the weights yourself, or only access it through an API?
▾
Open models (like Llama or Mistral) give you full control and privacy; closed models (like GPT, Claude, Gemini) trade that control for convenience and top-tier performance.
2🗜️
QuantizationShrinking a model's numbers to make it smaller and faster.
▾
Trades a small amount of accuracy for a big drop in memory and compute needs — this is what lets capable models run on a laptop or phone.
3📱
On-device AIRunning the model locally instead of in the cloud.
▾
Faster response and better privacy, since your data never leaves the device — usually a smaller, quantized model to fit local hardware.
4💰
Latency & cost per tokenHow fast a response comes back, and what it costs to generate.
▾
Bigger, smarter models are usually slower and pricier per token — picking the right model size for the job is a real, practical skill.
Module 3 · Don't Mix These Up
Commonly Confused Terms
Tap a card to see the difference.
🌀Hallucination vs. Bias
Both sound like "the AI got it wrong" — but they're different problems.
Hallucination is inventing a false detail. Bias is a skewed pattern inherited from training data. One is a random error; the other is systematic.
👆 tap for the difference
🔎RAG vs. Fine-tuning
Both "customize" a model — for very different reasons.
RAG gives the model new facts to reference. Fine-tuning changes how the model behaves or writes. Need facts? RAG. Need a new skill or style? Fine-tuning.
👆 tap for the difference
🌡️Temperature vs. Context window
Both are "settings," but control totally different things.
Temperature controls how varied the wording is. Context window controls how much information the model can hold in mind at once. One's about style, the other's about memory.
👆 tap for the difference
🎛️Token vs. Parameter
Both are numbers behind the scenes — but at very different levels.
A token is a chunk of your input/output text. A parameter is an internal learned value inside the model itself. Tokens flow through the model; parameters are the model.
👆 tap for the difference
Module 3 · Knowledge Check
Quick Check: Terminology
1. Which term describes a shared standard for connecting AI models to outside tools and data?
MCP (Model Context Protocol)
Temperature
Quantization
2. You want a smaller, faster version of a model to run on a laptop. What's this process called?
Fine-tuning
RAG
Quantization
Module 4 · Applying It
Choosing the Right Tool: Decision Flow
Click what you need to do:
🤖 Automate a multi-step task
📄 Answer questions from your own files
💻 Refactor a large codebase
🔗 Connect two apps together
✍️ Get more consistent output from a chatbot
👆 Choose a task above to see the recommended approach.
Module 4 · Live Activity
Practice: Build a RAG-Ready Prompt ⏱️ 5 Minutes
Check off each step as you complete it.
🎉 Nice work — that's the core pattern behind every RAG-powered tool you'll use.
Module 4 · Hands-On Practice
Common Intermediate Mistakes
Shuffle through the mistakes — one at a time.
🤖
Over-trusting agent autonomy
Letting an agent take high-stakes actions (sending, purchasing, deleting) without review.
💡 Give agents the least access they need, and review anything irreversible before it happens.
1 of 4
Module 4 · Cross-Cutting Topic
Using AI Responsibly at This Level
Tap a card to ask yourself the hard question.
🎣Prompt injection awareness
Content an agent reads (a web page, an email, a document) can contain hidden instructions trying to hijack it.
🤔 "Would I trust this source enough to let it control what my agent does next?" Treat retrieved content as data, not commands.
👆 tap to ask yourself
🔐Least-privilege permissions
Give agents and tools only the access they actually need for the task.
🤔 "Does this agent really need write access, or would read-only get the job done?" Default to less access.
👆 tap to ask yourself
🗃️Data governance
Know what data a RAG system or agent can see — and whether it should.
🤔 "If this document were leaked through an AI answer, would that be a problem?" Scope access accordingly.
👆 tap to ask yourself
🧾Verifying agent actions
Autonomy is convenient, but a clear audit trail matters once mistakes are possible.
🤔 "Could I explain exactly what this agent did and why, if someone asked?" If not, add logging or a review step.
👆 tap to ask yourself
Module 4 · Level Up
🏗️ Your Roadmap: Next Up, Advanced
1🔌
Building with APIs, hands-onMove from "understanding APIs" to actually calling one.
▾
You already know what an API is — Advanced is where you send your first real request and get a response back in code.
2🤖
Deploying agents end-to-endFrom understanding agents to running one that does real work.
▾
Configure permissions, tool access, and guardrails for an agent that actually operates in your workflow.
3🔎
Building a real RAG pipelineFrom the concept to a working system on your own data.
▾
Set up a vector database, chunk your documents, and connect retrieval to a live model.
4🎛️
Fine-tuning basicsFrom knowing what it is to trying it on a small example.
▾
Prepare a small training set and fine-tune a model for one specific, narrow task.
1 / 4 explored
🏅 Ready for Advanced!
Module 4 · Interactive
Find Your Focus
Click the goal that sounds most like you:
🔗 Automation-Minded
I want to connect AI into the tools and workflows my team already uses.
👨💻 Builder-Minded
I want to actually build with AI, not just use it through a chat window.
🔬 Researcher-Minded
I care most about accuracy, sourcing, and trustworthy AI answers.
🧭 Team-Enablement-Minded
I want to help my whole team adopt AI well and safely.
Module 4 · Recap
Recap & Resources
Click a card for where to go deeper.
🧰
The tool landscape
Revisit Module 1 any time you're choosing between a chatbot, an agent, and an automation platform for a task.
⚙️
How AI actually works
Tokens, embeddings, attention, context windows, RAG — the mental model behind everything else in this deck.
📖
The glossary
Bookmark Module 3 — it's the fastest way to look up a term the next time you hit one you don't recognize.
🏗️
This training's next level
Head back to the Main Menu and open Advanced when you're ready to start building.
Thank You
You Now Speak Fluent AI! 🎉
You understand the real toolbox of AI products, how models actually work under the hood, and the vocabulary that used to sound like jargon — plus where to go next.
✅
RecapSix tool categories · Tokens, embeddings & RAG · The essential glossary · Your roadmap to Advanced.
🙋
Questions?Now's a great time to ask, or revisit any slide using the navigation below.