Practise the interview behaviour, not only the answer.
A mock interview should test retrieval, clarification, structure, follow-up handling and communication under realistic uncertainty.
Behavioural warm-up
Tell me about a model that did not perform as expected.
STAR plus learning
Situation
Give only the context needed to understand the challenge.
Task
State your responsibility and the success criterion.
Action
Explain your decisions, trade-offs and personal contribution.
Result
Quantify the outcome honestly, including imperfect results.
Learning
Explain what you would repeat or change now.
STAR is not a script. It prevents a story from becoming vague, chronological narration. Spend most of the answer on your action and reasoning.
Worked STAR outline
| Part | Example |
|---|---|
| Situation | A churn model achieved good validation AUC but produced few useful retention actions. |
| Task | I owned the error analysis and threshold recommendation before launch. |
| Action | I checked calibration, segment performance and label windows; then worked with product to define intervention capacity and false-positive cost. We selected a threshold for the top actionable risk segment instead of maximizing accuracy. |
| Result | The launch served the retention team’s weekly capacity and improved contacted-customer conversion versus the previous rule-based list. |
| Learning | I would define the operational decision and intervention constraint before model training, not after evaluation. |
Reusable ChatGPT Voice mock prompt
Use only information you are comfortable sharing. Remove confidential company, customer and personal information before pasting material into any external tool. See the official ChatGPT Voice guide for setup and availability.
Make the mock harder deliberately
Change the constraint
Add latency, memory, privacy, label-delay or cost limits.
Challenge assumptions
“What if the classes become balanced next month?”
Demand evidence
“How do you know the improvement came from the model?”
Ask alternatives
“Why not use a simpler model?”
Inspect failure
“Which users or examples perform worst?”
Compress the answer
Repeat the same answer in two minutes, then thirty seconds.