Personalisation Engine and AOV Uplift
Dynamically served each user a tailored app layout based on their behavioural segment.
Blinkit's app served a highly diverse user base with very different purchasing behaviours — varying levels of price sensitivity, brand loyalty, appetite for new products, and average spend. The app was presenting a largely uniform experience to all users, leaving significant value on the table.
Build a personalisation system that could serve each user a customised app layout based on their individual behavioural profile, with the goal of increasing average order value.
The personalisation system had two interdependent components. The first was the segmentation model — a multi-dimensional behavioural classifier assigning each user a profile across price sensitivity, brand affinity, wealth index, and novelty propensity. The second was the delivery infrastructure: a reverse ETL pipeline loading segment outputs into a PostgreSQL database, served via an API that the mobile application called at each login.
The separation of these components was intentional — the segmentation model could be retrained and updated without touching the application layer, and new dimensions could be added without a mobile release cycle. The result was a system where every user's first action after login triggered a layout personalised to their behavioural profile, with each layer independently updatable.
- 01Analysed historical order data to identify the behavioural dimensions most independently predictive of order value and repeat purchase: price sensitivity, brand affinity, wealth index, novelty propensity.
- 02Built and calibrated the segmentation models: multi-dimensional clustering producing stable user-level segment assignments updated on a regular cadence from the warehouse.
- 03Implemented the reverse ETL pipeline: model outputs loaded from Snowflake into a PostgreSQL database structured for low-latency API reads.
- 04Built the segment API — called at user login to return the current segment assignments and feature flags for that user, consumed by the mobile app to determine layout configuration.
- 05Implemented LaunchDarkly to translate segment assignments into layout decisions — product ordering, featured placements, promotional surfaces — without hard-coded application logic.
- 06Designed and ran A/B tests to measure AOV uplift per segment configuration with statistical rigour; iterated on segment definitions and layout rules based on outcomes.
- Meaningful and sustained uplift of approximately 10% in average order value
- Personalisation system became a platform capability, extensible to new surfaces and signals over time
- Segmentation model and delivery infrastructure fully decoupled — model can be retrained and updated without a mobile release cycle
Start a conversation.
Every engagement begins with a focused discussion of your current data environment and priorities. To schedule an initial consultation, reach out directly.
Get in touch