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

End-to-end CTR prediction for push notifications

Click-through model, owned end to end, that decides which notification to send and when. +153% engagement.

What it is

A click-through-rate prediction model for push notifications, owned from research through production.

What it's for

For every notification slot, the system needs to decide which one to send and at what moment to send it, to maximize the chance the user actually taps. Owning the model end to end meant taking the project from problem definition to live production traffic.

How it was built

The full project lifecycle: data preparation, feature engineering, modeling, evaluation, training pipeline, online A/B testing, and production serving. Combined LightGBM and TensorFlow models depending on signal type. Lifted click-through rate by +153% over baseline.

My role

Sole author. Owned the full project lifecycle from research to production.

Built with
TensorFlowLightGBMscikit-learnHive SQLAirflowJenkinsGCP

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