Skip to main content

Intelligence

Space operations are entering a new era, and it will run on forecasts. Constellation is building the models behind them.

Why now​

Orbit is filling quickly. Every year brings more satellites, more operators, more ground stations, and more networks that have to share the same sky, spectrum, and time. The constellations already filed with regulators dwarf everything flying today.

Operations have not kept pace. Most fleets still run on static plans and alarms: a margin fixed months in advance, a threshold that fires after something has gone wrong, and a person on console working out why. That works when one team flies a handful of spacecraft. It breaks when one team flies hundreds, across networks it does not own.

The world those plans describe is not static. Weather moves over a ground station. A pass rises lower than planned. Traffic shifts with the hour and the day. Two objects drift toward the same point in space. Much of this can be seen coming, and what can be seen coming can be planned for.

What we are building​

Constellation is a research company building foundation models for space operations. We train them in house across categories: links and RF, network traffic, conjunction, weather, and contact planning. Each one learns from real telemetry and stays grounded in the physics that governs it.

Three principles shape every model we ship.

Forecast, then act. A model should tell you what will happen with enough lead time to change the outcome. An alarm only tells you what already broke.

Know what you do not know. Every forecast is a range, not a single number. We build intervals to be calibrated, so an 80% band is meant to hold the outcome 80% of the time. When a model lacks the inputs to answer, it declines instead of guessing.

Physics as the guardrail. Orbital mechanics, link budgets, and propagation are not things a model should relearn from scratch, or be allowed to violate. We build them in, and when measured telemetry disagrees with a model, the telemetry wins.

How operations change​

When forecasts are good and honest about their uncertainty, the operator's role changes. You stop reacting to each alarm and start supervising a fleet that looks ahead. You set the envelope: which actions are allowed, which limits matter, which decisions still need a person. Inside the envelope, systems act on forecasts. Outside it, they ask.

That only works if every action is accountable. Constellation is a control plane, so every forecast carries the model release and feature schema that produced it, and every decision traces back to the telemetry behind it. Autonomy becomes something you can audit, not something you take on faith.

Models today​

Two models are served now. Each page covers what to send, what comes back, and the limits of the current release.

  • Link SNR forecasts a link's signal-to-noise ratio 1 to 5 minutes ahead, with an 80% interval.
  • Traffic demand forecasts the traffic a pool will be offered over the next hour, in ten-minute intervals.

What we are training next​

  • Conjunction. Close-approach screening for your fleet: what is coming near, how close, and how sure we are.
  • Weather. Weather impact on links and passes: how the sky over each station changes what a link can carry.
  • Ground Segment / Contact Planning. Contact windows and station scheduling, with a plan that changes when the world does.

Build with us​

Every served model is available through GET /predictions and to AI agents over the MCP server. Start with the quick start, send your fleet's telemetry, and ask for a forecast.

Our research team works where physics, operations, and machine learning meet. If that is the work you want to do, write to contact@constellation.space.