Orbita

A solar storm can blind GPS. Orbita is an Alexa+ add-on that turns space weather into a go or a hold for this field, this machine and this job, and tells the crew when it clears.

Comment

Inspiration

On May 10, 2024, the strongest solar storm in twenty years hit the American Midwest in the middle of planting. Tractors stopped at the ends of their fields. Farmers reported GPS-guided tractors meant to hold a centimetre off by about ten feet and driving in circles, and had to stop until the storm passed. Kansas State University later put the losses to Midwest crop producers at an assumed $565 million.

NOAA had warned days ahead. But the warning said "Kp 9" and "G5". It did not answer the question a farmer actually has: can I plant this field, with this machine, today? And it says nothing at all on the quiet nights that still break a fix: in 2024 Hong Kong called off its National Day drone shows on two nights running because ionospheric scintillation over Victoria Harbour took away the drones' satellite positioning.

USDA puts auto-steer and guidance on more than half the US acres planted to corn, soybeans, winter wheat and cotton, and near the magnetic equator, where Brazil, India and much of Africa farm, the evening sky breaks satellite signals in season. I wanted the space weather feed translated into the farmer's own words, and into a decision.

What it does

Orbita is an Alexa+ add-on that turns space weather into a plain go or hold, for this field, this machine and this job.

  • "Can I plant tonight?" Near Grand Island, Nebraska, on a quiet night: go, with the expected drift in inches against the 2 inches planting needs. Near Sorriso, Brazil, on the same quiet night: hold off, because near the magnetic equator plasma bubbles after sunset break the RTK fix whatever Kp says. A Kp app calls that night clear.
  • "Watch it for me." Orbita watches the job and emails the crew when it clears, only after each crew member has agreed on their own phone.
  • "What happened to my planter on May 10, 2024?" It replays the real storm for that field: Kp 9 (G5), five of eight three-hour blocks outside planting tolerance, and the hours whose passes to check for guess rows.
  • Not only farms. Satellite teams ask whether the next three days will throw off their orbit predictions. A drone show is told when the sky clears. Survey crews hear about GPS jamming regions the sun can't explain. 23 kinds of operator get timed steps from the same data: satellites, grid control rooms, airlines, ships, undersea cables, data centres, small planes, households.
  • Machines ask too. Robots, drones and fleets get trust, verify or do_not_trust for up to 1,000 positions in one call, with signed alerts pushed to the fleet's own software.

On an Echo Show it draws three days of forecast against the job's own limit as an MCP App, with go and hold carried by position, hatching and words, never colour alone.

How I built it

The model talks; code decides. Every verdict and number is computed from NOAA, the measured sky and published sources, so no model can turn "hold off" into "probably fine". Three things sit between the data and the answer, and none of them is a model:

  • The job sets the tolerance: planting and strip-till 2 inches, spraying 20, tillage a metre, cadastral survey 3 centimetres, a drone show a decimetre.
  • The rig sets the baseline: RTK gives centimetres on a quiet day and fails first and hardest in a storm; WAAS starts coarse and degrades gently.
  • The site sets the exposure, by geomagnetic latitude, not the map: Winnipeg and Kyiv share a parallel and not a storm.

What is measured outranks what is forecast: the disturbance over the field in the last minutes, from about 4,000 satellite links (DLR's one-minute ROTI) and 27 magnetometers, can turn a forecast yes into a no. A source that could not be read is never counted as clear.

The one judgement in the model is how far poleward of NOAA's aurora view line the disturbed sky is taken to be overhead. It was chosen on 2024, against five-minute positions from real GPS stations at Saskatoon, Fargo, Grand Island and Lubbock, then tested on 2025, a year it had never seen: at Kp 6 and above it held 35 of 43 three-hour blocks where receivers really lost precise positioning, against 25 of 43 before calibration.

  • A self-hosted MCP server (spec 2025-11-25, Streamable HTTP, stateless, OAuth 2.1 with PKCE) in TypeScript on Node 20, with 18 tools, an MCP Apps panel and card, elicitation, and an Agent Skill readable as skill:// resources.
  • Answers from memory: NOAA SWPC (Kp, scales, alerts, solar wind, GloTEC, OVATION), DLR, USGS and INTERMAGNET, GFZ Potsdam's Kp archive since 1932, Coast Guard and FAA GPS notices and Open-Meteo are read on timers in the background, never inside a turn: p95 under 13 ms, against Alexa+'s 500 ms.
  • On AWS: EC2 behind an API Gateway HTTP API; DynamoDB for farms, one sealed (AES-256-GCM), versioned record per farm; the storm alert check as a Lambda run by EventBridge Scheduler every 15 minutes, alarmed in CloudWatch; SES for crew email; keys in SSM Parameter Store; all in CDK. An agent on the Strands Agents SDK follows each answer's next step and waits for a yes before anything is saved or sent.
  • The demo: while Alexa+ add-on access is limited to select partners, a Claude model plays Alexa+, reading the server's real tool list and calling the same live server.

Challenges I ran into

  • Kp is the wrong number on its own. The worst nights for a Brazilian farm are quiet ones. Plasma bubbles follow the sunset line along the magnetic field, so their season depends on longitude: October to March over South America, the equinoxes over Africa and Asia. An equinox-only rule said nothing to Mato Grosso in November, its worst month.
  • Half a second. Every NOAA feed is slower than an Alexa+ turn allows, so everything is fetched in the background and every answer is arithmetic from memory.
  • Proving it, not claiming it. A model that says "hold" too often is as useless as one that never does, so it was scored against real receivers and real years, not against itself.
  • The farmer's own words. "All my tractors are sitting at the ends of the field, auto steer won't engage, is it the sun?" became a test, along with speech-recognition noise, corrections, "never mind" and "what about tomorrow though".

Accomplishments that I'm proud of

  • Agrees with five-minute GPS station positions at Saskatoon, Fargo, Grand Island and Lubbock in 93 to 96 percent of 2024's blocks.
  • Fitted on 2024 and tested on 2025, it held 35 of 43 storm blocks where receivers lost precise positioning.
  • Matched 22 of 22 recorded operator actions in May 2024, November 2025 and January 2026.
  • 100 percent of 94 spoken sentences sent to the right tool, and 2,467 automated checks on every push.

What I learned

That the useful answer to a space weather question is almost never a number. It is a decision, a reason and a time: "hold off; plasma bubbles; it clears at 2 AM". That quiet days matter as much as storms near the equator. And that a safety answer has to say when it could not check, instead of defaulting to "clear".

What's next for Orbita

Real Alexa+ testing once add-on access opens, more correction networks and receiver models, push alerts to the Echo itself if add-ons gain a proactive channel, and working with equipment dealers and co-ops, who already field the calls when guidance fails.

The sun will storm again. Next time, everyone who steers by a satellite hears about it first.

Built With

Share this project:

Updates

Submission history