Ihza MahendraResume
Back to Production ML at scale
Production ML at scale

Push-notification audience targeting at 150M+ scale

Targeting model that predicts what 150M+ users want next and matches notifications to it.

What it is

A model that predicts which shopping category each user is most likely to want next, used to decide which push notification to send them.

What it's for

Sending the same notification to 150 million people is wasted reach. The platform needs to predict what each user is actually about to want and match the message to that prediction.

How it was built

Deep-learning sequence models that read user browsing history to predict their next category. Trained on Hive-warehoused behavior, served from a daily Airflow pipeline, and integrated into the production push system. Reached 0.80% conversion and 0.45% click-through across more than 150 million targeted users.

My role

Sole author, end to end.

Built with
TensorFlowDeep learningSequence modelsHive SQLAirflowGCP

Want the full technical depth, the tradeoffs, what broke, what I'd do differently? Ask the agent about this project.