Product Context
PayBridge serves small and medium-sized businesses through a payments dashboard and mobile app. The company wants to rank financing and payment-related offers—such as instant payout, working capital advances, card readers, invoicing tools, and cross-border payment discounts—so each merchant sees the most relevant offer at the right time.
Scale
| Signal | Value |
|---|
| Monthly active merchants | 8M |
| Daily active merchants | 1.6M |
| Peak offer-page QPS | 18K |
| Eligible offers in catalog | 12K total, ~200-800 eligible per merchant after policy filters |
| New / updated offers | ~1.5K per day |
| End-to-end p99 latency budget | 120ms |
Offer impressions happen on the home dashboard, checkout settings page, and email/push follow-up flows. Some offers have immediate feedback (click, apply), while others have delayed outcomes (approval, activation, repayment quality) that may arrive days later.
Task
Design an end-to-end ML system for personalized offer recommendation and ranking. Address the following:
- Clarify the product objective, target users, and success metrics for ranking SMB payment offers.
- Propose a multi-stage architecture for retrieval, ranking, and re-ranking, including what runs online vs batch.
- Define the training data, labels, features, and model choices for each stage, including how you handle delayed feedback and cold start.
- Design the serving stack with realistic latency and throughput assumptions, including feature storage, caching, and fallbacks.
- Explain how you would evaluate the system offline and online, and how you would monitor drift, training-serving skew, and business-policy violations.
- Identify key failure modes, especially around compliance, stale eligibility, and over-targeting merchants with low-quality offers.
Constraints
- Offers are subject to merchant eligibility, geography, underwriting, and compliance rules that must be enforced before ranking.
- Some labels are sparse and delayed by 7-30 days.
- The system must support new merchants with little history and new offers with limited interaction data.
- Cost matters: GPU-heavy online scoring is discouraged unless clearly justified.
- Explanations are needed for internal review on why a merchant received an offer.