A mini university for practical AI
Understand what makes an AI system useful beyond the first demo. Three focused lessons connect engineering decisions to examples you can inspect and exercises you can try.
Free access · no sign-up · learn at your own pace
What does a correct answer mean?
Where did the answer go wrong?
What happens when AI changes a system?
For developers, technical leaders and people choosing an AI solution for their team.
Your learning path
Follow the sequence or start with the problem you are facing. Each lesson includes an explanation, a worked example, a practical exercise and a question to check your understanding.
Quality before confidence
Turn a convincing answer into a testable result. Define reference questions, check source support and detect regressions.
A first evaluation set and clear acceptance criteria.
Evaluation · reference cases · regressions
Open the lessonFollow the evidence
Separate a retrieval failure from a generation failure. Inspect passages, document versions and the claims in the answer.
A repeatable way to diagnose an incorrect answer.
Retrieval · source support · freshness
Open the lessonFrom answer to action
Design tool contracts, permissions and recovery from partial failures. Make an agent's actions observable and reviewable.
A checklist for the first controlled integration.
Tools · permissions · recovery
Open the lessonUse a small example from your work as you read. The aim is to leave with a decision you can explain and a result you can check.
Interactive laboratory
Choose a question, change the retrieved documents and inspect what happens to the answer. Learn to separate a retrieval problem from an answer that misuses its sources.
All documents and company policies here are fictional. Questions, retrieval results and answers are predefined examples. This demo does not run a model, calculate embeddings or send your selections to an AI service.
Inspect relevance, scope and revision before trusting a passage.
Damaged deliveries
Record the order number and a photo of the damage. A damaged-delivery request does not require the customer to contact the carrier.
Section 3 — Damaged deliveries. Record the order number and a photo of the damage. A damaged-delivery request does not require the customer to contact the carrier. Forward the evidence to support; the support procedure defines the next step. This revision replaces the archived damaged-delivery policy.
Replacement requests
Create a replacement request for the support team to review. Include the order number and the photo of the damaged item.
Section 2 — Replacement requests. Create a replacement request for the support team to review. Include the order number and the photo of the damaged item. Creating the request starts a review; it does not automatically approve a replacement or guarantee a delivery date.
The example changes immediately when you select a question or evidence mode.
Supported by current evidence
Collect the order number and a photo of the damaged item.
Then create a replacement request for the support team to review. A request does not guarantee approval.
Diagnosis
Both passages are current and cover the question. Each action in the answer is supported by a cited passage, including the limit on automatic approval.
A citation identifies a source. It still needs to be current, relevant and strong enough to support the claim.
