Design recommendation and weapon-ad detection systems

You are asked to design two end-to-end ML systems.

1) Personalized recommendation system

Design a system that recommends items (e.g., videos/products/posts) to users in a large-scale consumer app.

Your design should address:

• Product goal: what you are optimizing for (e.g., engagement, watch time, conversions) and how you balance multiple objectives.
• Data: user events, item metadata, candidate sources, freshness, cold start.
• Modeling approach: candidate generation vs. ranking (and optional re-ranking), feature design, handling exploration.
• Serving: latency budget, online features, approximate nearest neighbor (if used), caching, fallbacks.
• Training & evaluation: offline metrics, online A/B testing metrics/guardrails, counterfactual concerns.
• Monitoring & iteration: drift, bias/fairness, abuse and feedback loops.

2) Weapon ad detection / ad safety classifier

Design an automated system that detects whether an ad (image/text/video/landing page) contains weapons or weapon-related content, and blocks or routes it for manual review.

Your design should address:

• Input modalities: text, image frames, video, OCR, landing-page text.
• Labels & taxonomy: what counts as “weapon content,” policy nuances, borderline cases.
• Modeling approach: multimodal models, thresholding, calibration, ensembling.
• Decisioning workflow: automated block vs. allow vs. send to human review; appeal process.
• Metrics: precision/recall tradeoffs, cost of false positives/negatives, SLA for review.
• Operations: active learning, handling adversarial ads, monitoring, model updates.

Assume millions of users/items for recommendations and high ad throughput (thousands to millions/day) for ad review. Clearly state any additional assumptions you need.