Architecture AI Updates: September 3, 2026

1. Adding a Second Task Made the Model Smaller and Better

Architecture. Swiggy’s in-house predicted lifetime value model is a multilayer perceptron with three shared hidden layers feeding task-specific heads for its Food and Instamart businesses, trained on more than 350 features spanning acquisition channel, device and fraud signals, geographic behavior, complaint history, category affinity, order behavior, payment patterns, and socioeconomic indicators. Every feature is available before a customer’s first order, which is what makes the model usable for cold-start ad bidding. The notable architectural result is the auxiliary task: predicting order count alongside lifetime value cut parameters from 363,000 to 135,000 — a 63% reduction — while improving accuracy, landing Spearman correlation above 0.75 for both business lines with 70-80% diagonal coverage across value deciles. The pLTV signal feeds Google’s target ROAS bidding in production. Source