Multi-domain recommendation platform using decoupled microservices with real-time inventory filtering, collaborative and content-based ranking, and manual curation overrides. Client requests flow through Kong/AWS API Gateway with A/B splitting to a Blender Service orchestrator that routes to specialized engines: Item Rec, User Rec, Accessories Rec, and Manual Rec, each powered by Python/FastAPI or Go. Real-time stop-list filtering via Redis/Dragonfly prevents out-of-stock recommendations, while NVIDIA Triton inference server ranks candidates using precomputed vectors from Milvus/Qdrant and live vectors streamed through Kafka/Flink. This architecture decouples recommendation logic from inventory, catalog enrichment, and ads services, enabling independent scaling and A/B testing. Fork this diagram to customize domain boundaries, add feature stores, or swap vector databases and stream processors for your infrastructure. Advanced practitioners can extend with feature engineering pipelines, real-time feedback loops, or multi-armed bandit allocation strategies.