Product Context

Intuit wants to personalize growth recommendations for a newly launched SMB segment inside QuickBooks, such as onboarding tips, financing offers, payroll nudges, and product education modules. The challenge is that this segment has little historical interaction data, but the experience still needs to feel relevant from day one.

Scale

Task

Design an end-to-end ML system to handle cold start for a new SMB segment in Intuit personalization surfaces.

1. Clarify the product objective and define what “good personalization” means for a new SMB with sparse or no behavioral history.
2. Propose a multi-stage architecture for candidate generation, ranking, and re-ranking across QuickBooks surfaces.
3. Explain how you would combine segment-level priors, content/item metadata, cross-segment transfer learning, and early-session signals to address cold start.
4. Define the offline training pipeline, online serving path, feature store design, and feedback loop while minimizing training-serving skew.
5. Specify evaluation, experimentation, monitoring, and rollback plans, including how you would detect feature drift and poor performance on the new SMB segment.

Constraints

• New segment has <14 days of historical data at launch and many businesses have zero prior interactions.
• Some labels are delayed (e.g., payroll activation, financing application, invoice usage) by days or weeks.
• Recommendations must respect compliance and eligibility rules for financial products.
• Cost target is <$0.001 per recommendation request, so heavy online deep models are limited.
• The system must support both batch personalization and real-time updates from fresh product interactions.