Building Disaster Pulse: What Happened When I Let AI Decide If a Disaster Is Real

When AI Becomes a Gatekeeper: Inside the Making of Disaster Pulse. Imagine living in a place where disaster is always one viral video away. In Indonesia, disaster alerts aren't just broadcast from official channels—they explode across TikTok, WhatsApp, news feeds, and desperate messages from people searching for loved ones. But not every scream online signals a real emergency. That's the chaos that inspired the creation of Disaster Pulse, an ambitious AI-powered platform designed to cut through the noise and decide, in real time, if a disaster is truly unfolding. At the heart of Disaster Pulse is a five-agent reasoning pipeline, a kind of digital relay race where each AI agent tackles a single, crucial task. First, the Observer sifts through raw signals—texts, videos, social posts—extracting objective observations. Next, the Classifier labels the event: Is it a flood, a fire, something else? How severe? The Skeptic, perhaps the most vital, challenges the assumptions, searching for errors or hallucinations and preventing embarrassing mistakes—like confusing a concert smoke machine for a raging inferno. The Synthesizer then weighs the evidence, scoring its confidence in what's happening, before the Action agent makes the call: raise an alert, update an incident, or ignore the noise. The platform's true innovation is its ability to analyze not only text but also video in the wild. Social media videos are dissected frame by frame, searching for floodwaters, fire glows, or collapsed buildings. And with a strict freshness rule, old footage—so often recirculated in moments of panic—gets flagged and downgraded, ensuring only current events trigger real-time alerts. But building a system this transparent required more than clever AI. Disaster Pulse had to show its work, not just its results. A disaster alert is only as trustworthy as the evidence behind it. That's why the platform features an AI Transparency Panel—a simple “Why?” button that reveals the entire reasoning chain, from raw signal to final verdict, giving disaster coordinators the confidence to act or wait. The journey wasn't without challenges. Multi-agent pipelines can fail when one agent's vague summary poisons the whole chain. Video analysis quickly becomes expensive, so clever frame sampling and caching were essential. And when demonstrating the platform for the first time, the empty state—no disasters, no alerts—nearly sabotaged everything until a demo data seeding system was built. Perhaps most striking was the moment the AI proved more cautious than its creator. While testing with what seemed like a dramatic, live news report, the Skeptic agent correctly flagged that the photos were from an old disaster. It was a humbling, trust-building revelation: Sometimes, the system is better at spotting mistakes than its builder. Disaster Pulse isn't just a technical marvel—it represents a shift in how communities might trust, scrutinize, and act on information in moments that truly matter. A system that's not just faster, but more careful, more transparent, and more reliable. In a world where misinformation travels as fast as disaster itself, giving AI the power—and the responsibility—to decide what's real could be the difference between chaos and coordinated response.