LIKELY QUESTIONS
- How have you built and deployed a fraud detection model from scratch in production, and what were the biggest trade-offs you had to manage?
- How would you improve an existing fraud risk model at Wise while keeping false positives low and protecting good customer experience?
- What features, labels, and evaluation metrics would you use for real-time fraud detection in international payments?
- Tell me about a time you identified a new fraud pattern or emerging threat from large-scale data and turned it into a practical mitigation.
- How do you approach model monitoring, drift detection, retraining, and incident response for fraud models in production?
- Wise mentions Python and collaboration with Java services. How have you worked across data science and engineering boundaries to productionize models reliably?
- How would you use unsupervised learning or anomaly detection in a fraud prevention setting where labels are delayed or incomplete?
- As a Lead Data Scientist, how would you influence investigators, product managers, engineers, and risk stakeholders when your recommendation creates short-term customer friction?

BEHAVIOURAL QUESTIONS
- Tell me about a time you had to balance fraud prevention performance with customer experience.
Model approach: Situation: fraud losses were rising but the existing rules/model blocked too many legitimate users. Task: reduce fraud while improving approval or pass-through rates for good customers. Action: segmented users by risk, analyzed error patterns, redesigned thresholds/features, partnered with product and ops, ran an A/B test, and added monitoring by customer cohort. Result: fraud loss reduced and false positives decreased, with a clear explanation of business impact and lessons learned.

- Describe a time you disagreed with engineers, risk, or product stakeholders on the right fraud solution.
Model approach: Situation: cross-functional team had conflicting priorities such as speed, model complexity, explainability, or operational burden. Task: align the group around an effective and practical solution. Action: brought data on expected fraud catch, precision, latency, operational cost, and customer friction; clarified constraints; proposed options and trade-offs; listened and adjusted; drove a decision with documented rationale. Result: team alignment, successful launch, and stronger trust across functions.

- Tell me about a time a model underperformed after deployment or a fraud attack bypassed controls.
Model approach: Situation: live model metrics deteriorated or a new attack vector emerged. Task: contain risk quickly, diagnose root cause, and restore performance. Action: used dashboards and slice analysis to isolate the issue, checked drift and label delays, introduced temporary safeguards, retrained or added features/rules, and communicated status to stakeholders. Result: losses stabilized, model performance recovered, and permanent monitoring or process improvements were implemented.

- Give an example of how you helped raise the technical bar or mentor others in data science.
Model approach: Situation: team needed stronger modeling standards, experimentation discipline, or better production practices. Task: improve team capability beyond individual delivery. Action: introduced review standards, reusable pipelines, documentation, metric definitions, mentoring, and regular knowledge sharing; coached others on stakeholder communication and production readiness. Result: faster delivery, fewer model issues in production, improved quality, and visible growth in team capability.

SMART QUESTIONS TO ASK
- How does Wise currently measure success for fraud models beyond pure detection rate, especially around customer friction, investigator efficiency, and downstream business impact?
- What are the biggest fraud threats or attack patterns the team expects to face over the next 12 to 18 months?
- How are responsibilities divided across data scientists, ML engineers, software engineers, analysts, and fraud investigators when taking a model from idea to production?
- What does strong performance look like for this role in the first 6 months, and what are the highest-priority problems the new hire would own first?
- How mature are the current model monitoring, experimentation, and feedback-loop systems, particularly for delayed labels, drift, and regulatory auditability?

RED FLAGS TO WATCH FOR
- They cannot clearly explain ownership, decision-making, or how data science works with engineering, fraud ops, and product; this may signal weak execution and constant stakeholder friction.
- They focus only on fraud catch rate and not on false positives, customer friction, explainability, or operational usability; this can indicate an unhealthy risk culture.
- They are vague about production infrastructure, monitoring, data quality, labeling, or model governance; this may mean you will inherit unreliable systems without support.