Product Context

PulseText is a cloud messaging platform used by banks, ride-share apps, and e-commerce companies to send transactional and promotional SMS worldwide. Design an ML-driven system that ingests, processes, prioritizes, and delivers millions of SMS messages per minute while maximizing delivery success and minimizing spam risk, carrier throttling, and user harm.

Scale

Task

1. Clarify the product objective and define what the ML system should optimize: delivery rate, latency, fraud/spam prevention, route selection, and prioritization under congestion.
2. Design an end-to-end multi-stage ML architecture for online message handling, including candidate route retrieval, route ranking, and final policy-based re-ranking or filtering.
3. Specify the offline and online data pipelines, labels, feature computation, and how you avoid training-serving skew across message, sender, recipient, and carrier features.
4. Propose model choices for each stage and explain tradeoffs between accuracy, latency, interpretability, and operational cost.
5. Define evaluation: offline metrics, online experiments, business guardrails, and segment-level analysis for high-risk senders, new routes, and international traffic.
6. Identify major failure modes at scale, including feature drift, feedback delay, carrier outages, abuse attacks, and stale routing models.

Constraints

• Some labels are delayed: final delivery receipts may arrive seconds to hours later depending on carrier.
• Compliance constraints vary by country; certain content classes or senders must be blocked or audited.
• The system must degrade safely: if ML is unavailable, transactional traffic must still flow.
• Cost matters: the platform cannot score every route with a heavyweight model for every message.
• New senders and new carrier routes create persistent cold-start problems.