Platform / Ops / Storm-Physics AI
Operations & Consumer

Storm-Physics AI
Models that are graded by the sky.

A research layer where every forecast is a testable claim: predictions are logged, checked against what actually happened, and only better models survive. Two headline results — a detector for the radar signature of debris a tornado is actually lofting, and a global model that flags hurricanes about to explode in strength — plus the discipline to say what isn't ready.

● Research — not yet in the consumer app Prediction loop runs on SynergyDB A model only ships if it beats the old one
What it replaces
Bespoke met-lab toolingone pipeline from raw data to proven accuracy Unverified model claimsevery number here was measured, not asserted
The flywheel

Predict → verify → retrain

The core is a loop, not a model. Every prediction the system makes is saved with a timestamp; when reality arrives, the prediction is graded against it; and a new model only replaces the old one if it scores better on the same fixed test — the system refuses a model that isn't better.

  • Predictions logged every cycle, graded automatically once the real outcome is known
  • Accuracy measured by how far ahead it looks — strong at 60 minutes out, fading fast after, and the system knows that about itself
  • A promote-only-if-better gate on every new model: verified never to get worse
  • Searches decades of hurricane history for storms that looked like today's — "what did similar storms do next?" — checked against known real-world cases

Measured, not asserted

0%false alarms on the hail cases that most often fool debris detection
~80%of violent tornadoes detected within effective radar range
~0.86skill score (AUC) — how often the model correctly ranks the hurricanes that are about to explode in strength, worldwide
The results

Two models worth talking about

🌪️Tornado-debris detector

Detects the radar signature of debris a tornado is actually lofting — physical confirmation a tornado is on the ground doing damage. On the hardest trap, hail that mimics that signature, it produced zero false alarms, while catching roughly four in five violent tornadoes within effective radar range. Telling real debris from look-alikes is exactly where working meteorologists struggle — this attacks it head-on.

🌀Global explosive-strengthening model

Predicts which hurricanes and cyclones are about to explode in strength — the forecast miss that kills. Trained on storms worldwide rather than Atlantic-only, it stays accurate in the Southern Hemisphere, where most published models go blind, correctly ranking the storms about to strengthen roughly 86% of the time (a skill score of about 0.86).

Both are research results on the shelf, deliberately: neither is wired into the consumer app yet. The prediction loop exists precisely so that when they ship, they ship with a measured track record instead of a marketing claim.

The direction

An AI meteorologist that only says what the data supports

The stated destination is an AI meteorologist that explains the weather in plain language. The design rule is absolute: it narrates over data the physics models already computed — radar, storm environment, proven accuracy scores — and never invents a forecast the models can't back.

Liability is treated as the binding constraint. This is life-safety information; a wrong confident sentence is worse than no sentence. That's why the strongest internal results stay internal until they clear their gates — and why "experimental, defer to official warnings" is stamped on everything.

Machine-learning modelsLevel-2 radar — the raw, full-detail feedNational radar + satellite feedsSynergyDB similarity searchPredict → verify → retrain loop
Keep exploring

Related