DEVOURED
Real-Time Personalisation at Scale: How Zepto Understands What You Want, Right Now
Zepto optimized real-time product recommendations by combining long-term user history with immediate session intent through a dual-transformer architecture.
What: Zepto implemented a 'Dual Sequence ReRanker' using separate transformer encoders for long-term history and short-term session actions. The system uses a 'target-aware pooling' mechanism to dynamically build user profiles against specific candidates, allowing for hyper-relevant ranking in milliseconds.
Why it matters: Standard static user embedding models fail to capture 'right now' intent. Fusing long-term behavioral data with real-time session signals is the current gold standard for improving conversion rates in high-frequency e-commerce.
Decoder
- ReRanker: A secondary machine learning model that takes a small set of top candidates from a retrieval system and orders them based on precise, complex features.
- Transformer Encoder: A neural network architecture that processes sequences of data (like clicks or page views) to understand their relationships and patterns.
Original article
Zepto built a Dual Sequence ReRanker for real-time personalisation that combines a user's long-term history with their current in-session behavior using separate transformer encoders for history and session sequences, target-aware pooling (dynamically rebuilding the user profile per candidate item), a learned fusion gate, and real-time signals (trending counters, calendar context).
Explore all 57 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi