Swiggy Deploys Multi‑Task MLP to Boost Customer Lifetime Value Forecasts

Swiggy has rolled out an in‑house predictive model that estimates each user’s lifetime value (pLTV) using more than 350 pre‑order signals and a multi‑task multilayer perceptron. By feeding the model an auxiliary task—order count—the company cut its parameter count by nearly two‑thirds while sharpening forecast accuracy, a move that directly powers its Google Target ROAS campaigns and promises higher margins for both its food‑delivery and Instamart divisions.

The engine rests on a dense MLP that ingests a rich tapestry of features: demographic tags, historical cuisine preferences, time‑of‑day ordering patterns, payment method usage, and even geo‑grid activity before a purchase is made. A secondary head predicts the raw number of orders a user is likely to place, allowing the shared hidden layers to learn a more generalized representation of customer behaviour. This multi‑task setup slashed the total trainable weights by about 63%, reducing GPU memory footprints and speeding up inference, while still delivering a mean absolute error improvement of roughly 8% over the previous single‑task baseline. The output feeds directly into Google’s Target ROAS bidding, enabling real‑time bid adjustments based on each shopper’s projected contribution.

Swiggy’s approach mirrors a broader shift in the Indian on‑demand sector, where rivals such as Zomato, Uber Eats, and BigBasket are also embedding AI‑driven LTV estimators into their ad‑spend engines. The Indian food‑delivery market, now exceeding ₹2 lakh crore annually, is increasingly dependent on micro‑targeted promotions to sustain growth amid intense price competition. Cloud‑native MLOps platforms from AWS, GCP, and Azure have made it feasible for mid‑size firms to train and deploy models of this scale, turning what was once a luxury for global players into a competitive necessity for domestic platforms.

For India’s tech ecosystem, Swiggy’s success validates a growing talent pipeline in data engineering and machine learning. Start‑ups building feature‑store solutions, such as HyperVerge and SigTuple, stand to gain from the demand for reusable pipelines that can handle hundreds of real‑time signals. Moreover, the model’s efficiency gains lower the barrier for smaller e‑commerce players to adopt predictive analytics without massive infrastructure spend, potentially democratizing advanced AI across the country’s burgeoning online retail landscape.

Key Highlights

• Launched a multi‑task MLP that predicts customer LTV with 350+ pre‑order features
• Auxiliary order‑count task trims model size by 63% while boosting accuracy
• Integrates pLTV scores into Google Target ROAS, improving ad‑spend ROI by double‑digit percentages
• Marketing and growth teams gain per‑user profitability insights for smarter budget allocation
• Future rollout plans include extending the model to hyperlocal logistics and opening an API for partner use

Real-World Impact

From day one, Swiggy’s data scientists, product managers, and performance marketers can act on granular LTV scores, allowing bid adjustments at the individual user level. Finance analysts will see tighter revenue forecasts, while ad‑operations teams reduce wasted spend on low‑value segments. The model also creates new demand for ML engineers skilled in multi‑task learning and real‑time feature pipelines, reshaping hiring priorities across the sector.

Why This Matters

The deployment signals a strategic pivot from reactive discounting to proactive value‑based acquisition. CTOs should prioritize building feature stores capable of ingesting hundreds of event streams and adopt multi‑task architectures that extract more signal per parameter. Developers will need to embed model inference into ad‑tech stacks, ensuring latency stays under milliseconds for real‑time bidding, thereby future‑proofing their platforms against a data‑first marketplace.

Swiggy’s next milestone is to fuse its LTV engine with the logistics layer that powers Instamart, aiming for end‑to‑end profit optimization. Watching how the company opens its predictive API to third‑party merchants will reveal whether this internal advantage can become a broader industry catalyst.