SYS//00 Mission Statement OPERATIONAL

Tornado detection, unplugged from the cloud.

TIMA Radar is an open-source AI that reads national weather radar the way an expert meteorologist does — spotting tornadic rotation signatures in real time. It's trained on 16,000+ historic tornado events and engineered to run on the hardware communities already own: no data center, no GPU fleet, no cloud dependency between a storm and the people in its path.

LIVE DETECTION ENGINE MRMS · 3-CH
TIMA Score · Grid Cell
0.000
TORNADIC SIGNATURE
AUTOMATED STORM TRIAGE RANKED BY THREAT
Cell IDCountyScoreStatus
TC-4471Ottawa, OK0.958Flagged
TC-4468Craig, OK0.641Monitor
TC-4465Labette, KS0.212Tracking
TC-4463Newton, MO0.087Tracking
SYSTEM TELEMETRY V8 · PRODUCTION
Training Index 0 events
Raw NEXRAD Corpus 0 TB
Inference Latency <100 ms / scan · CPU
RECRUIT//01 Citizen Science SPOTTERS WANTED

Can't code? If you can read a sky, you can train the AI.

An AI is only as sharp as the people who teach it. StormSwipe turns storm knowledge into training data: we deal you real archived radar frames, and you make the call — clear, severe, or tornado — in one tap. Think Tinder, but you're swiping on supercells.

Every verified label sharpens the detector — especially on its hardest problem, the rotating storm that looks tornadic but isn't. That's a judgment call trained spotters make better than any algorithm. If you're a SKYWARN spotter, a chaser, a dispatcher, or just the person your family calls when the sirens go off: this is your seat on the project.

No code required Real archived radar Streaks & leaderboards Every label trains v9
STORMSWIPE · TRAINING DEMO CARD 1 / 4
KTLX · ARCHIVED FRAME YOUR CALL, SPOTTER →
RESULT
Spotter Aptitude: High

4 / 4 correct. The network wants your eyes.

Join the Spotter Corps

Full app in development · early access opening soon

Labels: 0 Streak: 0 Accuracy:
MODULE//01 The Problem CRITICAL

Warnings measured in minutes. Decisions measured in seconds.

When a tornado forms, the national average warning lead time is roughly 13 minutes — and historically, a majority of tornado warnings verify as false alarms, eroding the public trust that makes people act when it counts.

The bottleneck isn't radar coverage. It's attention. During an outbreak, a handful of human forecasters must monitor dozens of rotating storms simultaneously, in real time, under extreme pressure. Signatures get missed. Triage happens by instinct.

And the emerging AI tools that could help are built cloud-first — useless to the rural county whose connectivity just went down with the same storm it's trying to survive.

~13 min
Avg. Warning Lead Time
National average from warning issuance to touchdown. Every added minute of confident detection saves lives.
Dozens
Storms per Forecaster
Outbreak-mode cognitive overload: more rotating cells than any human team can continuously watch.
0 bars
Cloud AI in a Disaster
Severe weather takes down the exact connectivity that cloud-dependent AI requires to function.
MODULE//02 The Edge-Compute AI Solution DEPLOYED

Train in the cloud once. Run anywhere, forever.

TIMA's detector is a deliberately compact convolutional network — about 242,000 parameters, thousands of times smaller than a chatbot, purpose-built for one job. Heavy training happens on rented GPU clusters; the finished model then scores a full radar grid box in under 100 milliseconds on ordinary CPU hardware.

That single architectural decision changes everything: near-zero marginal inference cost, no vendor lock-in, and detection that keeps running when the internet doesn't.

County EOC desktop Chase-vehicle laptop Fire-station mini-PC Edge server / SBC Data center required
DEPLOYMENT ARCHITECTURE 3-TIER
TIER 01
CLOUD
GPU Training — On Demand Batch training against the full event index and NEXRAD corpus. A bounded, one-time cost per model generation. Spend stops when training stops.
TIER 02
EDGE
Local Inference Node The 242k-parameter model runs on commodity CPUs or consumer GPUs. Sub-100ms per MRMS grid scan. Deployable air-gapped.
TIER 03
ANY BROWSER
Triage Console Lightweight ingest and API layer streams ranked threat scores, imagery, and telemetry to any workstation or phone.
MODULE//03 The Data COMPOUNDING

Trained on the largest storms in American history.

TIMA learned what danger looks like from the ground truth: significant tornado reports in the NCEI Storm Events database, matched to the raw radar volumes that captured them. Negatives aren't easy clear-air fillers — they include a deliberately mined set of strongly rotating storms that didn't produce tornadoes, forcing the model to learn the difference that matters.

And the dataset grows itself: every live scan of an active NWS warning polygon is automatically archived with its score, then verified against post-event ground truth. Routine operation is data acquisition — a proprietary training asset that compounds every storm season.

TRAINING CORPUS INDEXED
Tornado Event Index 0
Raw NEXRAD Radar Data 0 TB
Input Channels · MRMS 3 reflectivity + dual-layer shear
THE DATA FLYWHEEL AUTOMATED
01
ScanLive engine evaluates every active NWS warning polygon, every radar cycle.
02
ArchiveRaw input arrays stored alongside peak output score — automatically.
03
Verify & RetrainPost-event ground truth labels each case. Every season deepens the moat.

Honest Science, Published Openly

When early training produced a suspiciously perfect 0.9947 AUC, we diagnosed it as data leakage, rebuilt the protocol around strict leak-safe splits, and published the honest figure — 0.944 ROC-AUC on held-out data — along with our open problems. Full methodology, limitations, and false-alarm analysis in the technical report.

EVIDENCE//00 Field Validation LIVE & PUBLIC

Not claims. Receipts.

Since July 2026 a continuous tracking layer has watched every rotating storm across the MRMS domain — before and independent of official warnings. Each catch, each miss, and the running accuracy numbers are published from the unedited archives. Timestamps included; judge for yourself.

Causal replays — no hindsight Measured 3.2-min data latency stated Misses published beside the catches
CASE FILE 01 SASKATCHEWAN · 2026-07-10
Tracked at 90–99% before the official tornado warning 83 min

Interactive replay of the recorded track data for the Hafford/Lilac supercell, cross-verified against Canadian radar. Fragmentation and pending ground truth disclosed on the page.

Open Case File
CASE FILE 02 INDIANA · 2026-07-11
Single track at 77–85% before three NWS warnings 47–79 min

Radar-indicated tornado warnings for Posey County — the same rotation physics TIMA scores, flagged three-quarters of an hour earlier.

Open Case File
TRACKER SCORECARD REGENERATED DAILY

Warning anticipation, false-alarm bands, and triage-tier outcomes, recomputed every morning from the raw track and warning archives — including the numbers that aren't flattering yet.

View Live Scorecard
MODULE//04 Open Source & Public Safety AGPL-3.0

Built in the open.

TIMA Radar is released under AGPL-3.0. The architecture, training protocol, and evaluation methodology are published for the meteorological and open-source communities to inspect, reproduce, and improve. Life-safety AI should not be a black box.

It assists. It never replaces.

Built by a firefighter and trained storm spotter, TIMA follows one rule above all: it holds no warning authority, and a low score never guarantees safety. The National Weather Service is the sole official warning authority — always follow local emergency management.

MODULE//05 Call to Action

Fund the minutes that matter.

We are engaging angel investors, grant committees, and strategic partners to accelerate the roadmap: dual-polarization debris fusion, environmental late fusion, and temporal nowcasting — converting detection into lead time.


[email protected]