Product Context

AdPilot is a SaaS marketing platform for small and medium-sized businesses (SMBs). Its home dashboard recommends campaigns, audience segments, budget actions, and product tips; the challenge is launching personalization for a newly onboarded SMB segment with little or no historical interaction data.

Scale

The new segment has sparse behavior, limited CRM integrations at signup, and different business profiles from existing customers. You need to design an end-to-end ML system that serves useful recommendations on day 0, improves quickly as signals arrive, and avoids overfitting to larger enterprise segments.

Deliverables

1. Clarify the product objective, success metrics, and what “personalization” means for a new SMB with no history.
2. Design a multi-stage recommendation system (retrieval → ranking → re-ranking) that handles zero-shot and sparse-signal cases.
3. Define the offline and online data pipelines, including feature generation, labels, training cadence, and how to prevent training-serving skew.
4. Specify the serving architecture, latency budget by stage, fallback behavior, and capacity planning at peak QPS.
5. Propose an evaluation plan covering offline metrics, online experiments, segment-level analysis, and long-term monitoring.
6. Identify major failure modes for the new SMB segment, including feature drift, segment misclassification, and poor exploration.

Constraints

• Day-0 recommendations must work even if the business has only signup metadata (industry, region, company size, acquisition channel).
• Some integrations (payments, CRM, ads) arrive hours or days later, so feature availability is uneven.
• Cost target: serving compute under $0.0015 per request.
• Compliance: no use of sensitive attributes; all features must be auditable.
• Freshness: new campaigns and recommendations should be eligible within 15 minutes of creation.