Design a capstone demo that survives difficult questions

A capstone demo should make the project’s decision and evidence easier to inspect. A smooth interface is useful, but reviewers also need to see provenance, baseline, evaluation, failure and operating boundaries.

Allocate the demo around eight sections

The career evidence lab creates an authored 12-minute structure.

from career_cases import capstone_demo_case

result = capstone_demo_case()
assert result["minutes"] == 12
assert len(result["sections"]) == 8
assert len(result["hard_questions"]) == 6
assert result["complete"] is True

The sections cover the decision and user, data provenance, baseline, evaluation, one failure example, the workflow, operating limits and next decision. Every section maps to an evidence path. The check verifies structural completeness on a template; it is not a human assessment of a real presentation.

Show the model in a decision path

Start with one case and the action a user can take. Explain when the prediction is made and which data is available then. Show the current baseline before the candidate. Present the locked evaluation with denominators and a meaningful slice.

Demonstrate one normal input and one invalid or uncertain input. The latter should reveal schema validation, abstention, fallback or a useful error. If the service is only local, say so. A screen recording can protect the presentation from network failure, but keep the reproducible local command available.

Prepare for questions that challenge validity

The lab asks: What leaks? Why this baseline? Who bears a false positive? What failed? How is input validated? When do you roll back? Link each answer to an artifact rather than adding hidden slides full of claims.

State limits without burying them: synthetic data, delayed labels, small sample, missing external validation, local latency, pending human review or untested container build. Then describe the next evidence needed. Do not improvise a production or business-impact claim during questions.

Rehearse with deliberate interruptions. Give a reviewer the evidence index and ask them to choose the order. Time recovery from a failed demo input. Keep a versioned snapshot of outputs so live retraining cannot silently change the reported result.

The Data Science course aligns this demo with the proposed capstone rubric: framing, data and split integrity, evaluation, reproducibility/deployment, communication and limitations.

Exercise

Build the 12-minute run sheet and evidence map. Ask two reviewers to use all six hard questions plus one unexpected scenario. Revise only the presentation or missing artifact; do not tune the locked test result after the rehearsal.

Continue learning

This article is part of the Data science careers and portfolio decisions sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the preceding task in Decide whether to specialize in NLP, forecasting or tabular ML.
• Return to the cluster foundation in Choose a data science project from a business decision.

Reference: NIST AI RMF playbook.