Home labs & practical projects
Evaluate a small AI assistant against a task rubric
A project to do with your own tools, with guidance on what to create and capture.
What you could create
A support team wants an AI assistant to answer questions from a short knowledge base. It needs to judge usefulness and failure cases before adopting a service.
| Evidence | What to capture |
|---|---|
| Question set and evaluation rubric | Show the inputs, scope and decisions behind this record. Unanswerable questions are part of the evaluation. |
| Assistant suitability and risk assessment | Include the evidence supporting your conclusion. The assessment distinguishes fluent wording from supported answers. |
What you'll need
Document, spreadsheet or diagram tool
Write five fictional support articles and ten questions, including two that the articles cannot answer. Prepare a scoring rubric before reviewing example outputs.
Project record
Keep the working files for question set and evaluation rubric and record assumptions as you go.
Current reference material
Use the linked provider resources to verify the features, access terms and version relevant to your project.
Access & cost
No paid platform is needed for the written version. Use a document, spreadsheet or diagram tool you already have.
The organisation and records are fictional. Write down assumptions; do not substitute confidential employer information.
Suggested milestones
- 1
Plan and investigate
Define correct answers, acceptable uncertainty and evidence requirements.
- 2
Develop your approach
Evaluate responses for accuracy, grounding and appropriate refusal to guess.
- 3
Verify and explain
Identify data handling, human review and service-selection questions.
Useful resources
Use these resources to deepen your understanding and check key details.
A useful starting point
Write five fictional support articles and ten questions, including two that the articles cannot answer. Prepare a scoring rubric before reviewing example outputs.
What a good result looks like
- Unanswerable questions are part of the evaluation.
- The assessment distinguishes fluent wording from supported answers.
Finish the project
Keep the decision record and note unanswered questions. A written design is not evidence that a technical control has been tested.
Independent home-project guidance, not an official vendor lab or a claim of complete exam coverage. Use your own tools and label the work accurately.