Prepare a model project for a technical viva

A technical viva tests whether you understand the evidence behind the project. A polished interface cannot answer why a split was valid, which feature leaked, what the baseline achieved or whether the reported score can be reproduced.

Prepare a compact evidence pack before rehearsing explanations.

Build the defence pack

The career evidence lab defines seven artifacts and five core questions.

from career_cases import viva_case

result = viva_case()
assert result["artifacts"] == 7
assert result["questions_answerable"] == 5
assert result["ready"] is True
print(result["evidence"])

The pack contains a decision brief, data manifest, split manifest, metric record, error slices, model card and reproduction command. These filenames are an authored template; readiness still depends on their contents and reviewer inspection.

Trace each answer to an artifact

For “Why this split?”, open the manifest and show the prediction timeline or grouping key. For “What leaked in the rejected version?”, show the feature-availability table and invalid experiment. For “Which errors matter?”, connect a confusion count or ranked case to the user’s action. For “What would stop deployment?”, point to a predeclared gate. For “Can you reproduce it?”, run the documented command and identify the generated evidence.

Do not memorize one uninterrupted speech. Reviewers interrupt, change assumptions and ask for a specific record. Practise moving from claim to file to limitation in under a minute.

Rehearse adversarial questions

Ask a peer to challenge data rights, row grain, missing labels, test reuse, subgroup harm, latency, fallback and ownership. If the answer is unknown, say so and identify how you would test it. Inventing certainty under pressure weakens the entire project.

Keep a failure example ready. Explain why an attractive model or feature was rejected. Show what the evidence supports and where it stops—for example, a local synthetic benchmark cannot establish production latency or customer impact.

The Data Science course aligns capstone review with these artifacts so the viva evaluates reasoning and reproducibility alongside the demo.

Exercise

Create the seven-file pack for one project. Give a reviewer only the index and 15 minutes. Record every unanswered question, repair the evidence path and repeat with a different reviewer without changing the held-out result.

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 prerequisite or neighbouring task in Data science interview: interpret a failed experiment.
• Continue with Build a data science resume with verifiable project evidence.

Reference: NIST AI RMF playbook.