Scalable artificial intelligence (AI) workflows increasingly fail not because individual models are weak, but because the surrounding architecture cannot process heterogeneous, bursty, high-stakes evidence at operational speed. This paper proposes ECHO-DR, an Event-Centric Hierarchical Orchestration architecture for real-time disaster response. The realworld problem addressed is the difficulty of turning social media, remote sensing, UAV imagery, weather alerts, seismic feeds, and incident reports into timely, auditable, and trustworthy operational intelligence during floods, earthquakes, wildfires, and storms. ECHO-DR introduces four core contributions: an event-centric memory plane that unifies vector retrieval, geospatial indexing, lakehouse lineage, and structured event graphs; a hierarchical routing policy that escalates only high-value or uncertain items to expensive multimodal reasoning; a stage-disaggregated serving design that independently scales encoders, prefill workers, decoders, and tool calls; and a governance plane that embeds auditability, human review, and zero-trust access control into the workflow. A formal utility-constrained routing model, event-linking algorithm, fusion rule, and capacity model are developed to show how the architecture scales under large workflows. The paper also provides an implementation blueprint, clean system diagrams, benchmarking methodology, ablations, and simulated evaluation results. Simulated tracedriven experiments indicate that the proposed gated architecture can reduce p95 provisional alert latency relative to a monolithic multimodal pipeline while maintaining evidence traceability and limiting deep-model cost. The work demonstrates that scalable AI for future big workflows should be designed as a compound, event-centered, policy-aware system rather than as a single model endpoint.