Interview Prep

Key talking points for Club Piscine:

• $3.14M marketing spend with no channel-level measurement -> now optimized with 29.4% revenue lift
• Manual e-commerce reporting across 6 domains -> automated Power BI dashboards
• No demand forecasting for 287 store-division groups -> 21.2% wMAPE (vs 47.2% naive baseline)
• Franchise network needs inventory planning support for seasonal business

2. Knowledge of Project Details (25%)

What they're testing: Can you accurately describe data sources, architecture, modeling, UI, and explain decisions?

Data sources: 68 Power BI slice exports, Open-Meteo weather API (5 cities), python-holidays (QC+ON), Table_Magasins.xlsx Architecture: Azure Container Apps, medallion architecture (bronze/silver/gold), automated pipelines Modeling decisions to defend:

• Why LightGBM over XGBoost/Prophet for UC4? (Global model handles 287 groups, MAE objective, early stopping)
• Why direct revenue model instead of units x price? (Price-per-unit swings with product mix, produces NaN)
• Why Bayesian MMM for UC1? (Handles multicollinearity in media channels, probabilistic framework)
• Why single global model instead of tiered? (Deprecated approach only covered 5.9% of groups)
• Why walk-forward CV? (Respects temporal ordering, tests across all seasons)

Assumptions to state:

• Weather data from 5 representative cities covers 27 store locations
• 125 weeks of history sufficient for seasonal patterns
• Store regroupments (CP77->CP07, CP100->CP10) don't distort patterns
• Lag warmup period (4 weeks) is adequate

3. Capstone Connection to Business Context (25%)

What they're testing: Can you link features to specific KPIs and business objectives?

Pain point -> Solution mapping:

4. Overall Understanding and Individual Contribution (25%)

What they're testing: Do you own your work? Can you claim at least one complex element?

Team Member Profiles (for targeted prep)

Practice Question Bank

Warm-up Questions

1. In your own words, what does Club Piscine do and why did they need this project?
2. Walk me through the three solutions at a high level.
3. What is your role on the team and what did you personally build?

Technical Deep-Dives

4. Explain how the demand forecasting model works. What features does it use and why?
5. Why did you choose LightGBM over other algorithms? What alternatives did you test?
6. How do you handle the fact that some store-division combinations have very low volume?
7. Walk me through the feature engineering pipeline. Which features matter most and why?
8. How does the weather data integration work? Why those 5 cities?
9. Explain the walk-forward cross-validation approach. Why not a simple train/test split?
10. What are the confidence intervals telling us? How are they generated?

Architecture Questions

11. Walk me through the data flow from raw data to final forecast.
12. Why Azure Container Apps? What alternatives did you consider?
13. How would this scale if Club Piscine opens 10 more stores?
14. What happens when the model needs retraining?

Business Value Questions

15. If you were the CFO, would you trust these forecasts? Why or why not?
16. What's the ROI of implementing this forecasting system?
17. How does the 29.4% marketing revenue lift translate to dollars?
18. Which stakeholder benefits most from which solution?

Tough/Curveball Questions

19. What didn't work? What approaches did you try and abandon?
20. If you had 6 more months, what would you improve?
21. Your winter wMAPE is 26.4% -- is that good enough for inventory decisions?
22. The CI coverage is 86.9% vs your 90% target -- what does that mean?
23. How do you handle model drift after deployment?
24. What happens if the weather API goes down?

How to Use This Skill

When invoked with a team member's name:

1. Read CLAUDE.md and relevant notebooks to ensure technical accuracy
2. Generate a personalized set of 10-15 practice questions based on their role
3. For each question, provide a model answer that references specific numbers, decisions, and files
4. Flag any areas where the member's knowledge might be thin
5. Suggest 3 "power moves" -- impressive technical details they can proactively mention

When invoked with a topic:

1. Generate deep-dive questions on that topic
2. Provide comprehensive answers with specific references
3. Anticipate follow-up questions the professor might ask