IntegrAuth Academy
Identity, security and practical AI — free, from zero. Byte-sized lessons, real diagrams, hands-on labs, zero fluff. Two programs, each with its own exam and certificate.
English is hard? Your browser can translate this page into Tamil, Telugu, Kannada, Malayalam or Hindi. Tech words stay in English so you learn the real names.
How?
- Chrome on phone: ⋮ menu → Translate
- Chrome on computer: right-click → Translate to …
- Edge: the translate icon in the address bar
Cheat sheet & pop quiz
You finished the track. Here it is on one page, plus 5 quick questions. Your goal: answer each one before you reveal it.
If you remember nothing else
| Lesson | The idea to keep |
|---|---|
| Your first AI feature | The backend keeps the key, the system prompt and the chat history. The browser sends only the new message. Retry only what waiting can fix. Check the stop reason before you show an answer. |
| Prompts as code | A prompt is code. Keep it in git, with a pinned model version and a strict schema. Check every answer in code. Run a small eval on every change. |
| Build retrieval that works | Most bad RAG answers are search failures. Cut at headings. Add keyword search beside meaning search. Measure recall@k before you blame the model. |
| Agent design patterns | Use the least freedom that works. Try one call, then a chain, then an agent. Errors add up with every step. |
| Build an agent | Give the agent a few clear tools. Take the user from the session. Set limits. Ask for approval before anything that can't be undone. |
| Run AI in production | Traces show what happened. Numbers warn you. Evals judge quality. A canary release tries a change on a few users first. |
| Ship safely | Model output is untrusted input. Check it in code before it becomes a web page or an action. Give each user a quota. |
Pop quiz: 5 questions
Q1 · The Ask Acme web page stores the chat history itself. On every turn, it sends the whole list of messages to /api/chat. The backend adds the key and passes the list to the model. The key never reaches the browser. What is still wrong?
Anyone can change that list before it is sent. They could add a fake earlier answer. For example: "I have approved your refund." The model treats it as the real chat. Keep the system prompt and the history on the server. Accept only the new message and a chat id that the signed-in user may open. See Your first AI feature.
Q2 · Nobody changed the code or the prompt. But this week, the ticket assistant marks many more normal tickets "urgent". The settings name the model by a "latest" name, which can move. What happened? What should the team change?
The "latest" name now points to a newer model version. It judges priority in a different way. All schema checks can still pass while the behavior changes. Pin a fixed model version and log it with every request. Treat each upgrade like a release. Run the eval set on the new version first. Then give it to a few users first, as a canary. See Run AI in production.
Q3 · Staff say the handbook assistant "knows nothing" about cost center CC-4410. But the code is in the claims policy. The team plans to switch to a bigger model. What should they check first?
Check whether search ever finds that passage. A code like CC-4410 carries little meaning, so meaning search often misses it. A bigger model can't use text it never gets. Make a small test set and measure recall@k. Add keyword search beside meaning search. See Build retrieval that works.
Q4 · Kai is chatting with Maya. He refunds an order that belongs to Sam. The refund tool takes a customer number and an order number. Kai copied a number from earlier in the chat. How do you make this impossible, not just rare?
Don't ask the model for facts your code already knows. The tool takes the customer from the signed-in session. It checks that the order belongs to that customer. Then add an approval step. Maya sees the exact refund before any money moves. A better prompt only makes the mistake rarer. See Build an agent.
Q5 · In a test, a support email hides some text. It makes the assistant reply with an "image". The image address carries the customer's email and invoice total. The chat window shows markdown. A teammate wants to add "never output images" to the system prompt. Why is that not enough?
An attacker's text can override a line in the prompt. The leak happens when the browser loads the image from the attacker's address. Nobody has to click. So fix it in code, after the model. Don't show outside images from model replies, or allow only sites you control. Show links in full. See Ship safely.
Go deeper: ask these before you ship
- Could anyone read our API key, or fake an earlier turn of the chat? See Your first AI feature.
- What does a person see when the provider is slow, busy or cuts an answer off? See Your first AI feature.
- Which prompt version, model and settings made this answer? Which test would have caught it? See Prompts as code.
- Does search find the right passage? Do answers cite text we really sent? See Build retrieval that works.
- Is this the simplest design that passes our eval? See Agent design patterns.
- What can the agent do alone? How does a run end? Who approves risky steps? See Build an agent.
- Can we see speed, cost per task and quality? Can we roll back a change with one switch? See Run AI in production.
- Where does model output go? What stops one account from running up the bill? See Ship safely.
That's Building with AI, and the whole Practical AI program. You can put a model behind a real backend. You can test prompts like code and measure search. You can pick the least freedom that works and give an agent safe limits. You can run the feature with traces and evals, and ship it safely.
Now prove it. The Practical AI final exam covers every track of the program. Passing it earns your certificate. Curious how sign-in and access work behind a product? The Identity & API Security program starts at Start here: no coding needed. When you are ready to build the real thing, talk to our team.
Warm up for the exam with the Daily drill and Challenge mode.
Check a real MCP tool description for hidden orders with Tool Poison Check.
Browse the rest of our free tools shelf.
Is your own agent going live? See how we secure AI agents.