Ground segment · Advisory only
VELA turns a telemetry deviation into a cited root-cause narrative and a drafted recovery procedure — every claim traceable to the spacecraft's own manuals and knowledge graph1, and held from release until an operator signs off.
Battery cell degradation — string 3 undervoltage
Deviation follows no related telecommand; cell-imbalance signature matches the FMECA degradation mode for the EPS battery.
The loop
VELA sits in the ground segment and does the investigation a mission-assurance engineer would do — faster, and with its working shown at every step.
A deviation is flagged — by onboard FDIR, or by VELA's own detectors over the downlink. Command-response noise is filtered out first.
A deterministic agent graph walks the anomaly, reaching data only through an audited boundary, and writes a root-cause narrative where every claim carries a citation.
The drafted recovery pauses at a real release gate. Nothing leaves VELA until an operator reviews the evidence and signs off. VELA never executes it.
Why teams trust it on a live fleet
Six properties that make an AI safe to put in front of a spacecraft operator.
Every narrative and procedure claim resolves to the spacecraft's own manuals and graph — or it is emitted as UNGROUNDED and held from release. A hallucinated reference is a sev-1.
VELA reads telemetry and drafts recovery steps for a human to authorize. It never constructs, formats, or transmits a telecommand — enforced by a grep guardrail in CI.
Runs fully offline with in-memory stores, or against your own Neo4j and Qdrant. No external calls, no CDN, no data leaving the ground segment.
Telemetry enters through one swappable source — CCSDS / PUS, XTCE, Yamcs, OpenC3, MQTT, or seeded replay. The reasoning above it never changes.
Graph traversal, event forming, and citation checks are deterministic and unit-tested; replay is seeded, so an investigation reproduces exactly — every time.
The drafted procedure pauses at a real approval interrupt. The operator's decision — and the evidence behind it — is captured in an append-only audit trail.
For the room the fleet depends on
Mission-assurance teams, insurers, and defence programmes can trust an AI on a live fleet precisely because it advises and cites — it never touches the commanding chain. That boundary is the product.
The boundary is the moat
Most tools bolt an AI onto the command path and ask you to trust it. VELA does the opposite: it stays on the read-only side, shows its evidence, and hands a human the pen. That is what lets it near a real spacecraft.
Plug into every tier — one swappable telemetry source
See it investigate
A guided walkthrough on the VELA-REF-1 reference spacecraft — no live mission data required. Run it in your own air-gapped environment.
Project