AutoML-Enabled Infrastructure for Adaptive Personalization in High-Traffic Shopping Events

The necessity for highly adaptable, ultra-low latency personalization systems is critical in high-traffic e-commerce environments, particularly during unpredictable demand surges such as high-volume shopping events. Traditional machine learning (ML) paradigms, constrained by manual feature engineering, model design, and slow iteration cycles, are fundamentally unsuitable for managing massive, unpredictable load and dynamically changing user intent during peak traffic events. This paper introduces a resilient architectural framework built upon the Automated Machine Learning (AutoML) paradigm, specifically optimized for deep recommender systems (AutoRecSys). The proposed architecture details an ultra-low latency feature pipeline integrating Apache Kafka for high-frequency ingestion, Apache Flink for stream processing, and dedicated Feature Stores (Redis/Delta Lake) to mitigate catastrophic train/serve skew and ensure comprehensive data consistency. The efficacy of AutoRecSys techniques such as sampling-based searches and multi-fidelity optimization is substantiated by demonstrating quantified production benefits, including architectural efficiency gains that result in up to a 25% Query per Second (QPS) increase and resource optimization that reduces embedding parameters by 50% to 95%. The holistic, integrated system consistently achieves an end-to-end latency below the critical 100-millisecond threshold, providing a robust, evidence-based blueprint for deploying adaptive, scalable ML solutions in high-stakes, time-constrained production environments.