Design a keystroke-by-keystroke address autofill system that ranks candidates based on partial text and user context

• Meet a strict 50ms latency budget across retrieval, feature assembly, and ranking stages
• Learn from implicit click labels while correcting for severe position bias in user feedback
• Personalize results using near-real-time behavioral signals without violating hard geographic relevance constraints

• FST prefix index with geohash partitioning
• LambdaMART with inverse propensity weighting
• Nearline behavioral feature updates