Roughly 12,000 satellites are active in LEO, and operations still run on human expertise. The bottleneck isn't fleet size — it's heterogeneity. Every mission has its own hardware, software, and quirks, and the burden lands hardest on small operators and university programs staffing ground stations around the clock.
Telemetry is the only witness.
Radiation and solar activity induce faults you never observe directly — only through telemetry. It's high-dimensional, mode-dependent, and welded to mission-specific procedures, so reading it well takes years of training that walks out the door with every departing operator.
LSTM finds it. RAG explains it.
We propose a ground-based operator agent with a hybrid architecture: LSTM models learn nominal and anomalous patterns from historical telemetry, while retrieval-augmented generation grounds each detection in mission documentation, procedures, and prior operational data. The agent flags the anomaly, explains it in the mission's own terms, and recommends remediation.
The hard parts are the point.
Turning noisy telemetry into trustworthy recommendations means guarding against outdated procedures, hallucinated commands, and unsafe sequences — enforcing deterministic safety constraints, respecting temporal ordering, and teaching the agent when to escalate or stay silent rather than act. Operator trust is earned in exactly those moments.
Train your own operator.
The goal: any organization can train a mission-specific operator agent from the telemetry and documents it already has, then keep refining it with operational data over time — from pre-launch rehearsal through years on orbit. Ground operations that scale like software, not like headcount.
MORE All insights →