Building an End-to-end ML Platform
Understand the components of an end-to-end ML platform, from feature engineering to real-time serving. We'll highlight the benefits of a centralized platform in accelerating model development cycles and improving reliability.
David Edison
Lead Machine Learning Platform Engineer at Dave
Brendan Hu
Sr. Manager, ML Platform Engineering at Dave
Building and maintaining machine learning models in production presents significant Engineering and Data challenges. In this talk, we will share lessons learned from developing a centralized, end-to-end ML platform designed to accelerate development and ensure reliability. We will walk through the core pillars of our strategy and the principles behind our architecture, focusing on how we address common failure points in the ML lifecycle. Key points will include:
- Establishing a "single source of truth" for features to eliminate training-inference skew and ensure data consistency between research and production environments.
- Accelerating the path from model ideation to deployment by creating a streamlined, low-friction workflow for data scientists that abstracts away underlying infrastructure.
- Delivering scalable, low-latency model inference that integrates seamlessly with feature data to power real-time product decisions.
- Implementing robust post-deployment observability for system health, model performance and data drift.