This lesson starts the capstone project of the AI track. No new libraries are added. What's new is the question every lesson asks: How does someone who has never seen your code tell that an AI feature is finished? With ordinary code, a glance and a test are often enough. With an AI feature, the model answers a little differently every time. "Works on my machine" therefore proves nothing. At the end of this module, the Kladde proves each of its AI features itself, with tests, test sets, attacks and numbers.
The module has six lessons:
| Lesson | What happens | Hours |
|---|---|---|
| T11.1 Planning | inventory, issues with a test set, definition of done, costs, risks | 2 |
| T11.2 Building I | assessment and support chat finished, with acceptance tests | 4 |
| T11.3 Building II | quote agent with approval and MCP server with permissions | 4 |
| T11.4 Operations | evaluation in the build, traces, monthly cap, README | 3 |
| T11.5 The attack test | at least ten attacks, a report, the attacks in the build | 4 |
| T11.6 Wrap-up | demo, code reading and technical discussion | 3 |
This module has no module exam. The documented attack test and the technical discussion in T11.6 take its place.
What the Kladde can do at the end
The Kladde still belongs to the invented small IT office. At the end of the AI track, it has six things that together make up its result:
- an assessment of tenders with a structured answer, with prompt version and model on every result,
- a support chat that answers only from its own help pages, names its sources and refuses without a source,
- a quote agent that drafts quotes and needs an approval before sending,
- an MCP server for the Kladde's data, with permissions per tool,
- evaluation in the build for each of these features,
- a monthly cap for the AI costs.
On top of that come a documented attack test against your own agent and a technical discussion about it. You have already built almost all of it in T1 to T10, lesson by lesson. In this module, you bring it together and finish it.