ReliefQueue AI: Human-Review Disaster Coordination

ReliefQueue AI helps disaster-response teams turn fragmented reports into a safer, reviewable coordination workflow. Field teams can capture and update cases through an offline-aware interface, local coordinators can manage affected zones and operational context, and command-center operators can review priorities, assignments, messaging, queue resilience, audit evidence, and AI advisories from one product. The application is deliberately human-in-the-loop. AI can summarize reports, expose missing information, identify duplicate or contradictory evidence, suggest operational tags, and support prioritization, but it never claims confirmed rescue, guaranteed safety, or automatic dispatch. Every AI output remains review-required. The AMD evidence campaign used AMD Developer Cloud with an AMD Instinct MI300X and vLLM 0.23.0+rocm723. Across a staged composite corpus of 24 single-report, complex-dossier, and adversarial cases, the campaign resolved 24/24 cases with 100% normalized JSON, nonce binding, and source coverage. The average measured generation throughput was 225.975 completion tokens/second. The evidence is labelled honestly: it was a staged calibration campaign rather than one uniform production-prompt run, and the direct-endpoint campaign did not exercise the application fallback path. ReliefQueue also demonstrates deterministic degraded operation, local queue recovery, dead-letter review and replay, role-scoped interfaces, public redaction, audit trails, and deployment through a single Python process serving the built React application and Product API.