AI-Driven Personalized E-Commerce Recommendations
Ranking and retention models to boost conversions and customer retention.
Key Details
| Challenge | Homepage merchandising was static; conversion lagged on return visits. |
|---|---|
| Solution | A ranking service using catalog, session and purchase signals. |
| Technologies | Python, Redis, BigQuery, TensorFlow Recommenders |
Technologies used
Python
TensorFlow
Redis
Google Cloud
Docker
Client background
A fashion and lifestyle retailer ran static homepage merchandising. Return visitors saw the same grid; conversion and repeat purchase lagged peers who personalized by session and purchase history.
Key challenges
- Merchandising rules could not keep up with catalog and season changes.
- Session signals were unused beyond basic “recently viewed”.
- No feature store — models could not go live without brittle batch jobs.
- Product teams lacked online metrics tying ranking changes to conversion.
What we built
- Ranking service for homepage and PDP modules using catalog + session features.
- BigQuery pipelines and a Redis-backed feature cache for low-latency serving.
- TensorFlow Recommenders models with offline eval and online A/B hooks.
- Ops dashboards for conversion lift, coverage and cold-start handling.
Project team: 6 engineers across AI/ML, backend and domain specialists — delivery over 15 weeks.
How we delivered
01
Baseline
Measured current conversion and identified high-traffic modules.
02
Features
Built session, catalog and purchase features with freshness SLAs.
03
Serve
Deployed ranking APIs behind feature flags for controlled rollout.
04
Learn
Ran A/B tests and fed winners back into merchandising defaults.
Business impact
LiftOn conversion
BetterRepeat purchase
LiveFeature store