Product Context
PayBridge is a fintech platform used by small and medium businesses to manage payments, working capital, and merchant services. When an SMB owner opens the dashboard or applies for financing, the platform must rank the most relevant payment offers such as card processing plans, instant payout, invoice financing, BNPL, and business credit products.
Scale
| Signal | Value |
|---|
| DAU | 4.5M SMB accounts |
| Peak QPS (offer ranking requests) | 18K |
| Offer catalog | 12K active offers across issuers, geographies, and segments |
| Candidate pool per request | ~500 eligible offers before ranking |
| p99 latency budget | 120ms end-to-end |
| New / updated offers | ~1,500 per day |
| Training events | ~220M impressions and 9M conversions per day |
Task
Design an end-to-end ML system to rank payment offers for SMBs. Your design should address:
- How you would define the objective, labels, and success metrics for offer ranking
- A multi-stage architecture for eligibility filtering, retrieval, ranking, and optional re-ranking
- The data and feature pipelines, including online vs batch features and feature freshness requirements
- Model choices for each stage, including how you would handle cold-start users, sparse conversion labels, and delayed outcomes
- Offline and online evaluation, experimentation, and launch strategy
- Monitoring, failure modes, and how you would detect feature drift and training-serving skew
Constraints
- Compliance: offers must respect geography, underwriting rules, lender eligibility, and adverse-action constraints
- Latency: p99 must stay under 120ms, including policy checks
- Cost: average serving cost must remain below $0.0015 per request
- Data availability: some labels are delayed by days or weeks (e.g., funded loan, retained merchant volume)
- Explainability: business teams need reason codes for why an offer was shown or suppressed
- Freshness: newly launched offers should be eligible within 15 minutes, and user payment behavior features should update within 5 minutes
