About the service
I set up and maintain ML feature store on Feast so your models see the same features during training and in real time. Training/serving skew usually creeps in unnoticed: an offline pipeline and online service compute the same feature slightly differently, and it surfaces later in degraded metrics when diagnosing is costly and slow. I eliminate this entire class of errors — each feature is defined once and served from a single source of truth.
The work includes the full cycle: entity and feature definitions, offline store for historical data to train on, online store for low-latency serving, and point-in-time correct joins so training data doesn't leak from the future. I set up load pipelines, materialization schedules, and monitoring for feature staleness and drift. Everything is version-controlled and fully reproducible.
You get a clean feature registry reused across projects, a faster path from experiment to production, and matching numbers on both sides. I can start from scratch or fix a half-built store, and I document each step in detail so you can maintain it afterward.
The work includes the full cycle: entity and feature definitions, offline store for historical data to train on, online store for low-latency serving, and point-in-time correct joins so training data doesn't leak from the future. I set up load pipelines, materialization schedules, and monitoring for feature staleness and drift. Everything is version-controlled and fully reproducible.
You get a clean feature registry reused across projects, a faster path from experiment to production, and matching numbers on both sides. I can start from scratch or fix a half-built store, and I document each step in detail so you can maintain it afterward.
