Deploying a Multistage Multimodal Recommender System on Amazon Elastic Kubernetes Service

This document details the design and deployment of a multistage, multimodal recommender system on Amazon Elastic Kubernetes Service (EKS). The system architecture consists of four stages: a Two-Tower retrieval model, a Bloom filter to exclude recent interactions, a DLRM ranker for scoring, and a final reranking step. It leverages a comprehensive MLOps stack including NVIDIA Merlin for training, Triton Inference Server for serving, Kubeflow for pipeline orchestration, and Feast for feature management. The models incorporate both collaborative signals and multimodal content features from CLIP and Sentence-BERT embeddings to handle cold-start scenarios and provide context-aware recommendations. The end-to-end pipeline covers everything from data preprocessing and model training to scalable serving and continual fine-tuning.