Your goal: show a plain account of why models make things up, a costed set of defenses, and
the judgment to pick a setting by risk.
How this works (about 30 minutes)
- Felix asks which role you are interviewing for (product, engineering or other), then one
question, and goes deep. Ask him about the product, its users and what a wrong answer costs;
he hands over what he has, one line at a time. The details are made up for the exercise.
- Follow-ups, one curveball, then a trade-off turn: your fix improves one number and worsens
another. Ship, hold or change?
- If you want to sketch or jot, say "can I use the notepad?"
- No hints and no scores during the call. Your report comes after.
What scores well
- Questions about the product and the cost of a wrong answer before any fix
- A plain one-line account of why it sounds sure, then the two kinds of failure
- The cause for this product named before the cure
- Defenses with their costs: ground, cite, abstain, verify, constrain, a person
- A way to find failures and to measure wrong answers beside coverage
- A call that splits by risk, with a line that would reverse it
Tips
- Take Gen AI Coach · Hallucinations first.
- Ask Felix for the failure log.