One operator, 25 satellites, one night shift. Designing an AI teammate that can be trusted.
A concept for on-the-loop anomaly resolution: automation detects, diagnoses and plans, and a single operator supervises. The interface shows the mission rather than the machines, hides what doesn't need a human, and proves what it's hiding.
Why a concept piece
For the past six years I've designed mission-critical software for space operations: constellation management, space object tracking, commanding, mission management and AI-assisted anomaly resolution. That work can't be shown. This piece rebuilds how I approach those problems from public principles, with a fictional commercial operator, invented data and no real program's workflows. I set the problem, the principles and the scenario, directed the design and reviewed every screen; an AI assistant did much of the production.
Constellations are always connected now. People can't be.
Older operations waited for a ground pass, and the next contact window set the tempo. Satellites linked by optical crosslinks are in contact continuously, so automation can act in seconds. The operator's job shifts from planning contacts to supervising decisions made at machine speed, across many vehicles, alone, for a whole shift.
Abstraction you can always peel back.
From 14 alarms and a 37-step checklist to one situation and one decision.
A crosslink starts failing while the operator is on a break.
RAD-07's optical terminal is drifting off target as it heats up. Every screen below is playable; open any of them, or play the demo from the start.
-
{{b.n}}
{{b.title}} {{b.body}}Open
Small choices that decide whether people trust the system.
Can someone with minimal training make the right call?
The hypothesis is that this design cuts training time. It hasn't been tested yet. The plan is a small study with people who have no space background, running the same anomaly on the legacy console and on the new design.
| Measure | How |
|---|---|
| Time to the right decision | From the briefing to approving the plan, legacy console vs. new design |
| Errors | Wrong approvals, missed twist, approving without opening details |
| Situation awareness | Freeze-and-ask questions: which satellite, what is the default, how long is left |
| Trust calibration | Do participants catch a planted wrong automatic action in the ledger? |
| Time to competence | Minutes of instruction before a first unassisted run succeeds |