Looper: An End-to-End ML Platform for Product Decisions

By Igor Markov et al

Published on Aug. 14, 2022

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. ML for Smart Strategies
3. The Looper Platform

Summary

Looper is a general-purpose vertical ML platform developed by Igor Markov and his team. It covers the end-to-end ML lifecycle, from collecting training data to deployment and inference, with a focus on accommodating product engineers without ML backgrounds, supporting fine-grain product-metric evaluation, and optimizing for product goals. The platform introduces general principles and architecture for decision-making and feedback collection, extending support to personalization, causal evaluation, and Bayesian tuning. During the 2021 production deployment, Looper hosted a large number of ML models and made millions of real-time decisions per second. The platform aims to lower the crossover point for adopting smart strategies and delivering product impact across diverse applications. Modeling approaches include supervised learning, contextual bandits, and reinforcement learning. The platform extends end-to-end ML systems by integrating software-centric ML, enabling high-quality data collection and holistic experimentation. Additional requirements for smart strategies include handling metadata features and monitoring resource usage for efficiency. The Looper platform is designed with operational safety in mind, supporting lightweight models that can be retrained and deployed quickly on shared infrastructure. It separates application code from platform code and leverages existing horizontal ML platforms for interchangeable models.