Steriod adds a second, explainable opinion behind your mission-control system: rules your engineers approved, each citing the operating document it came from, plus a label-free anomaly detector benchmarked on ESA's own public dataset. Every alarm names the affected channels, fires once per event, and traces back to the number and the document behind it.
Every commanded manoeuvre, eclipse and mode change trips a threshold. Operators learn to ignore alarms.
A score without a reason does not survive a review board, and it cannot be told apart from a commanded event.
A single event fragments into dozens of alarms across channels. Someone has to stitch them together by hand.
Root cause means custom queries over raw archives, timelines rebuilt by hand, and the one engineer who remembers the last time.
Items marked Pilot support are delivered with the pilot, configured for your mission by our team.
| Deliverable | What it is |
|---|---|
| Approved monitoring rules | Readable rules drafted from your operating documents, flight procedures, ICDs and mission database, approved one by one by your engineer. Each cites its source. |
| Label-free anomaly detection | Operating envelopes, seasonal baselines, slope and joint-behaviour models mined from your own archive, with no training labels. Each mined bound carries its derivation and sample count, and stays unarmed until a human approves it. |
| One alarm per event, with affected channels | Rule failures fold into episodes on the mission-time axis. Each episode lists the channels involved and whether a telecommand preceded it. |
| Command context, never silent suppression | Command-to-effect links are mined from your archive. A rare nominal event caused by a command is annotated with that command; it is suppressed only after a person accepts the link. |
| Root-cause investigation | On request, one correlated investigation across every failing rule: time windows, charts and statistics computed by the engine, a narrative written by an LLM bound to cite those numbers, confidence stated. "Undetermined" is an allowed verdict. |
| Panels inside your console | The rule board and episode panels embed in Yamcs, Grafana or your operations console as chrome-less pages. |
| Operational qualification sheetPilot support | How the system meets the operator requirements published with the ESA benchmark (binary verdicts, affected channels, alarm once for rare nominal events, online streaming, single high-end PC), with per-rule alarm statistics, reproducible replays and an exportable audit trail. |
| Fleet viewPilot support | The same rule set on every satellite of the same bus; the unit that deviates from its siblings is ranked first. |
Every screen below runs on ESA's public Anomaly Detection Benchmark data (ESA/ESOC, Airbus DS, KP Labs, CC-BY 4.0).
Envelope rules on three channels, a joint model across six, coupling and noise-band detectors, each with its peak against its threshold. Two mined telecommand sequences account for every span, so the episode is annotated, and nothing is suppressed until a person accepts the link.
A learning window on the mission clock runs the miners: envelopes, joint models, coupling, noise bands and command-to-response maps. Every result is staged for review; a named person closes the learning phase.
For any signal: what moves before it, what follows it, and whether the documents predicted that relation. Couplings measured on recorded runs but "not predicted by the system graph" are findings worth a look.
The engine seeds the investigation with the failing rule, the exact time windows and the requirement behind it. An agent selects the signals and charts within a fixed budget; deterministic steps execute; the summary may only cite what was computed. Without grounding, the run says so instead of guessing.
A read-only tap on your telemetry archive, or a subscription to the live parameter stream. Signal metadata from your mission database.
A learning window on the archive, on the telemetry's own clock. Envelopes, seasonal bounds and command-to-effect relations are mined. Nothing leaves the node.
Your engineer reviews mined bounds and drafted rules next to their derivation and source, and arms the ones that should alarm.
The deterministic engine evaluates every approved rule on the live feed. Episodes appear on a shared timeline and inside your console.
Press "Start investigation". The engine assembles the facts, the model proposes a cause within the evidence, your engineer accepts or rejects.
Monthly: alarm statistics per rule, drift re-fits proposed with their evidence, rules retired or tightened by your team.
| You run | How Steriod connects | Stays untouched |
|---|---|---|
| SCOS-2000 / EGS-CC | On-premises node inside the control-centre network; archive replay first (read-only), live tap second; signal metadata from the MIB Pilot support | Your MCS, alarms, procedures and flight dynamics |
| Yamcs | XTCE mission database imported directly; parameters forwarded by Steriod's Yamcs-to-MQTT connector; Steriod panels in a Yamcs web tab | Your Yamcs instance and its alarm definitions |
| Commercial or in-house mission control | Your stack publishes decoded parameters to Kafka or MQTT, or drops JSONL archives | Your software, your command path |
| Grafana or an operations console | Rule board and episode panels embed as chrome-less pages | Your dashboards |
| Analysis archives (Terma STAT, data lakes) | Read from file exports Pilot support; the detector runs on the data you already keep | Your archive |
| Your own agents or scripts | Everything reachable as REST, command-line and MCP tools; deliberately no "approve" tool | Your automation |
Deployment: one Ubuntu server or equivalent, Docker Compose, offline bundle, no cloud dependency. LLM steps (document import, rule drafting, investigation narrative) run on an open-source model on your hardware; detection, evaluation and recording need no LLM at all.
Self-reported on ESA's public labels, scored with ESA's own code, walk-forward protocol. Full case study →
| Lever | Today | With Steriod | Basis |
|---|---|---|---|
| False-alarm triage | About 15 false alarms per satellite per month at about 2 h each: 4,320 h a year for 12 satellites | Triage volume cut by about 90 % through command context and seasonal baselines: about 350 k USD a year | Steriod ROI model, illustrative (12 LEO smallsats, 90 USD/h) |
| Root-cause investigation | About 80 engineering hours per major anomaly, 12 a year | About 4 h of human review over an evidence-bound verdict: about 82 k USD a year | Steriod ROI model, illustrative |
| Archive queries during investigations | Custom queries over raw logs, hours | Tens of milliseconds over a 14-year span | Measured |
| Week | What happens |
|---|---|
| 1 | Node installed in your network; archive connected; signal catalogue imported and reconciled |
| 2 | Learning window run on the archive; mined bounds reviewed; first approval session |
| 3 | Blind replay over the archive; episodes compared with your known anomalies; alarm statistics |
| 4 | Rules drafted from operating documents; coverage report; command context tuned |
| 5 | Live tap or shadow mode; two investigations run on real episodes |
| 6 | Qualification sheet, statistics and the approved rule set handed over; decision on operational use |
Priced as a fixed pilot fee, quoted on scope. The rules, bounds and evidence stay with you.
Blind replay first, scored against your own anomaly list afterwards. Or book a meeting with us at Space Tech Expo Europe in Bremen, 17–19 November 2026.