ML for Fraud Detection with Feast and Flyte
Details
The features your model trains on need to match the features it serves on, whether they're computed in a nightly batch job or fetched in real time at inference. A feature store closes that gap. In fintech, where a missed signal costs real money and a regulator will eventually ask exactly how your model produced a decision, that consistency isn't optional.
In this hands-on workshop, we'll build an end-to-end fraud detection pipeline with Feast as the feature store and Flyte 2 as the orchestrator. We'll engineer point-in-time correct features and train a model in a Flyte workflow, then deploy the model and Feast online store as a Union-hosted app for real-time predictions. Every input, output, and model is a versioned artifact with full lineage, and the same code scales from the workshop's local setup to production.
What we'll cover
- How feature stores keep batch training and real-time serving in sync
- Building point-in-time correct datasets with Feast
- Orchestrating training with Flyte 2: cached data prep, durable runs, and full lineage from source to model
- Deploying the model and feature store as a Union app for low-latency inference
- The path from demo to production: streaming ingest, drift monitoring, and scaled serving
What you'll leave with
- A working fraud detection model served behind a real-time API
- A reusable Flyte 2 pipeline you can adapt to your own data
- A portfolio-ready project you can adapt to a production scenario at work
Who it's for ML engineers and practitioners working on fraud, risk, or any production ML problem in a regulated environment.
Hosted by Sage Elliott, AI Engineer at [Union.ai](https://union.ai/?utm_source=luma)