Misdeclared hazardous cargo inside sealed maritime containers continues to drive ship fires, port disruptions, and avoidable losses for crews, carriers, and coastal communities. Ports and carriers already use non-intrusive inspection (NII) imaging and document checks, but these systems are often treated as separate queues rather than a single integrated decision system. This paper proposes a practical multimodal framework that fuses radiographic sensing with shipping document intelligence to flag hazardous misdeclaration risks without opening a container. The approach combines a vision encoder that learns density-aware patterns from X-ray or gamma imagery with a language model that extracts and normalizes claims from bills of lading, manifests, and booking descriptions. A fusion layer learns cross-modal consistency, so the system can react when what the scan suggests does not match what the paperwork claims. The design is grounded in port constraints, including strict throughput targets, noisy and delayed labels, long-tailed hazard categories, and an operational need for clear explanations that can be audited. We define a data lifecycle that turns inspections, holds, claims, and incident investigations into structured feedback, without requiring constant full unpacking of cargo. We describe a low-latency edge deployment pattern that reduces backhaul and helps avoid unnecessary centralized compute, which matters in regions where water and power constrain expansion planning. A simulation-driven evaluation plan is provided, including realistic cost-sensitive metrics that focus on recall at fixed false-alarm rates, because ports pay real costs for every extra secondary inspection. The paper positions the framework relative to DHS work on AI-enabled paradigms for non-intrusive screening and recent industry adoption of AI screening for dangerous goods in booking workflows.