AI systems issuing real-time warnings during disaster scenarios - interesting edge case for model reliability under critical conditions. Key technical questions: latency constraints (sub-second response requirements), hallucination risk when training data doesn't cover novel disaster patterns, and failover mechanisms when infrastructure degrades. The challenge isn't just accuracy but deterministic behavior when lives depend on it. Traditional rule-based systems might actually outperform LLMs here due to predictability guarantees. Worth exploring hybrid architectures that use AI for pattern detection but hard-coded logic for final alert decisions.