Where Steriod stands — and where it's going
We build Steriod in the open with a small number of early teams, so it seems only fair to write down, plainly, what the platform can do today and what we're working on next. This is the first post in what will become a regular status series.
The short version
Steriod is an AI-assisted verification & validation platform for satellite telemetry. Written requirements become Axiom rules — plain-text, human-approved, machine-executed. Around that core, three tracks have matured furthest this year: document intelligence, trend analytics, and the agent platform.
What works today
Document intelligence
Reading the customer's specification is the platform's strategic first track, because everything downstream — drafted rules, coverage numbers, root-cause citations — inherits its quality. Today the platform imports a specification PDF into a knowledge base with one structured artefact per requirement, searchable by meaning. On top of that sit coverage and gap analysis (which spec clauses do your test suites actually cover?) and drafting candidate rules for the gaps.
Two things we're particularly glad we invested in: imports run as durable background workers — resumable, cancellable, fully logged — and import quality is measured, not assumed. A verification loop scores every imported artefact against the source document, so a knowledge base carries its own trust score.
Trend analytics
Telemetry from live pipelines and recorded runs lands in a TimescaleDB trend store. You can ask questions about it in plain language, and the design here reflects a lesson we learned early: LLMs plan; engines compute. The AI plans in small, bounded, auditable steps — each one verified against the actual plan document before it counts. Execution is fully deterministic against the trend store, and the summary you get afterwards is grounded in the computed results. This staged flow turned out to be reliable even with small, locally hosted models — which matters for teams that can't send telemetry to a hosted API.
Multi-run comparison, run-scoped root-cause timelines, and durable, chainable analysis results round out the track.
The agent platform
All AI features run on a shared backend agent runtime: durable records, replayable event streams, pause, resume, and cancel. The Vi assistant runs server-side with durable conversations and approval-gated write tools; long jobs like document import and bulk rule drafting are handed off to background worker agents; and everything agents produce lands in a persistent per-project workspace that both humans and future agent sessions can read.
One principle has survived every redesign: approval is human-only. No agent in the system — including agents you connect over MCP — has an approve capability. AI drafts; humans sign.
The rest of the platform
- Continuous requirement V&V with four-state rule status (pass / fail / pending / not-exercised), live anomaly detection on MQTT streams, and YAMCS integration.
- The Axiom Playground — the full compiler and runtime in WebAssembly, in your browser.
- A Cesium-based 3D mission view with live, replay, and demo modes.
- A satellite simulator with scriptable anomaly injection for end-to-end rehearsal.
- The whole stack deployable with Docker Compose, and an MCP server so your own agents can work against the platform.
What we're building now
- Document intelligence: extraction-quality improvements driven by the measurement loop's findings, faster ingestion through provider-aware parallelism, and configurable embedding models.
- Analytics: chained analyses (one step consuming a previous result), compute-only steps, and snappier planning turns.
- Agents: server-side attachment handling so document flows are fully backend-resident, and finer cancel/approval controls.
- Platform hardening: resilience work on the trend store's recording path and the storage backends — the unglamorous track that makes the rest dependable.
- Desktop: bringing the standalone desktop app up to parity with the server deployment.
Further out
Three directions we're excited about, in various stages of design:
- Onboard execution. Axiom already has an embedded C runtime; the goal is the same signed rule text running from the test bench to flight-adjacent hardware.
- Requirement reverse-engineering. Learning candidate rules from historical telemetry to bring legacy systems back under an explicit, reviewable spec.
- More verticals and sources. The document pipeline generalises: requirement tools (Jira, Polarion, PTC/Jama, Confluence) as import connectors, and automotive data (ARXML, CAN traces) as a second domain.
Talk to us
If any of this maps to a problem your team has — or you think we're wrong about something — we'd genuinely like to hear it: ping@steriod.ai. And if you want the two-minute version of what Axiom feels like, the Playground is open.
Published 2026-08-11 · The Steriod team