python3 -m http.server 8000
✕
Between 2020 and 2026, Israel was the target of an unprecedented sustained campaign of rocket, missile, and drone attacks from Hamas in Gaza, Hezbollah in Lebanon, the Houthis in Yemen, and the Islamic Republic of Iran. Under Fire is an attempt to structure, visualise, and preserve that record.
At its core, Under Fire is an experiment — an exploration of whether frontier AI models can be used to aggregate, structure, and visualise a large, fast-moving dataset drawn from multiple, often conflicting perspectives.
The Iran–Israel conflict generates data at a pace that overwhelms traditional OSINT methods. Salvos arrive in rapid succession. Munitions counts vary across sources. Casualty figures are revised. Launch sites are identified days after impact. Israeli, Iranian, American, and international outlets offer conflicting accounts of the same events.
This project asks a simple question: can a combination of frontier AI, open-source intelligence, and human editorial oversight produce a structured, geo-visualised record of the conflict that is accurate, transparent, and useful? The answer so far is a qualified yes — and the site you're reading is the result.
The dataset is a six-year sweep of every alert recorded across Israel between January 2020 and March 2026 — every siren, every shelter, every 90-second sprint. Each event is attributed, where the evidence allows, to one of four actors: Hamas, Hezbollah, the Houthis, or Iran.
The underlying alert feed is sourced from RocketAlert.live, which relays Israel's Home Front Command warnings in real time. Origin attribution is estimated from the geography of the impact area and the timing of the salvo — cross-referenced with public IDF, IRGC, and press reporting. Where attribution is uncertain, alerts are tagged as unknown rather than guessed.
The pipeline behind Under Fire combines large language models with a spread of open-source intelligence sources, anchored throughout by human editorial oversight. None of these three legs would be sufficient on its own.
Anthropic's frontier model handles research, source synthesis, data structuring, and the bulk of the site's development. A project of this scale and pace would not be possible without modern frontier models.
RocketAlert.live as the primary alert feed. IDF and IRGC public statements. Reuters, AP, Al Jazeera, Times of Israel. Satellite imagery from Planet Labs and Maxar. Research from ISW, CSIS, and the Alma Center.
AI handles synthesis and structuring. Humans provide editorial judgement, source verification, and the contextual understanding that models cannot yet replicate — especially in a contested information environment.
Tens of thousands of munitions. Arrow, David's Sling, Iron Dome, THAAD, Patriot, Aegis — a layered defence architecture that intercepted the vast majority of incoming threats. Behind each of those interceptions is a name, a shift, an engineering decision made years before, and a ninety-second warning that someone had to act on.
This is a dataset that deserves to be structured, preserved, and made accessible — both as a historical record of what was endured, and as a demonstration of what open-source intelligence, combined with modern AI tools, can achieve.
This project is dedicated to every soldier who stands between Israel and the storm of rockets, missiles, and drones recorded on these pages — to the aerospace engineers who designed the interceptors, the radar operators and intelligence officers who track every launch, the pilots and missile operators who fire them — most of whose names we will never know.
And to the brave Iranians working from within Iran to bring about a peaceful country — one that does not threaten its neighbours, but grows together with them.