Design in-session personalization at scale

Design a real-time personalization system that adapts within a single browsing session: as a user clicks/scrolls/adds-to-cart over a few minutes, their recommendations should reflect that just-now intent (e.g. they switched from browsing shoes to looking for a gift), not just their long-term history. This must work for 100M users with sub-100ms recommendation latency. Design how in-session signals are captured, turned into features, and incorporated into ranking fast enough to feel live, and the trade-off vs batch-trained models.