RETHINKING MLOPS: BUILDING MODULAR, SERVERLESS MACHINE LEARNING PIPELINES ON AWS

Abstract

The rapid adoption of machine learning (ML) across industries has been hindered by the complexity, inefficiency, and compliance limitations of conventional container-based MLOps frameworks. This paper introduces a novel, cloud-agnostic, serverless MLOps architecture that integrates three key innovations: (i) privacy-by-design mechanisms using distributed differential privacy with adaptive budget allocation, (ii) cost-aware optimization through reinforcement learning–guided resource scheduling and fine-grained cost attribution, and (iii) AI-powered developer automation for natural language pipeline specification and predictive maintenance. We present both theoretical foundations and a reference implementation built on AWS, while maintaining portability across providers. Empirical evaluations on diverse real-world datasets—including e-commerce churn prediction, recommendation systems, and financial fraud detection—demonstrate that our framework reduces total cost of ownership by 35–65%, sustains accuracy within 2–3% of non-private baselines, lowers deployment time from days to hours, and improves fault recovery by up to 41%. These findings establish serverless MLOps as a viable alternative to traditional architectures, offering scalable, cost-efficient, and trustworthy ML deployment. The framework provides a pathway toward sustainable enterprise adoption of ML, with implications for compliance, developer productivity, and multi-cloud resilience.