WorkSeptember 2, 2026via InfoQ AI/ML
Swiggy Uses 350+ Features and Multi-Task MLP to Predict Customer Lifetime Value
Why it matters
A food-delivery giant's practical deployment of multi-task learning to predict customer value and optimize ad spend reveals how practitioners are moving beyond single-task models to compress complexity while improving performance. This is the applied-ML story: not a new model or capability, but how existing techniques solve real business problems.
Key signals
- 350+ pre-order features engineered for pLTV prediction
- Multi-task MLP architecture with order count as auxiliary task
- 63% reduction in model parameters while improving predictive performance
- pLTV signal integrated with Google Target ROAS for customer acquisition optimization
- Deployed across both Food and Instamart product lines
- Swiggy (food-delivery/e-commerce at scale)
- 350+ pre-order features in the pLTV model
- Multi-task MLP architecture (order count as auxiliary task)
- 63% reduction in model parameters while improving performance
- Deployed with Google Target ROAS bidding for customer acquisition
- Applied across Food and Instamart verticals
- Published Sep 2026 on InfoQ (engineering-focused outlet)
The hook
350+ features. 63% fewer parameters. Swiggy's lifetime-value model shows how real companies are deploying ML at scale—not in the lab, but in acquisition.
Swiggy developed an in house predicted lifetime value model using more than 350 pre order features and a multi task MLP for Food and Instamart. Adding order count as an auxiliary task reduced model parameters by 63% while improving predictive performance. The pLTV signal is used with Google Target R…