Network segmentation
The edge node sits in a Purdue Level-2/3 DMZ. Outbound-only to the control plane, no inbound sessions, no lateral reach into the plant network.
Security & trust
We write to your mill. So we earn it.
Steelira is an OT system with an AI inside, not an AI product with an OT bolt-on. Segmentation, clamps, watchdogs, deterministic fallback and immutable audit are architecture, not features.
SOC 2 Type II [ASPIRATIONAL]IEC 62443 alignedGrade-IP never cross-trained
request F5 gap -0.018mm signer=agent:roll-gauge
clamp envelope [-0.05, 0.05] ok
rate 3 writes / 10s limit 6 ok
grade DX54D policy approval > 0.03mm ok
watchdog last heartbeat 41ms ok
write level2.setpoint ack=true audit=7f31c
Reviewed by OT, quality and security teams at design-partner sites
NordStrip Steel
Kalyon Metals
Ferrata Group
Aurum Flat Rolled
Meridian Coil
Volkan Çelik
Architecture
How the control path is protected
Six controls, all enforced in the runtime rather than in policy documents.
Signed, clamped writes
Every setpoint request is signed by the emitting agent and clamped per setpoint, per grade and per line before it reaches your Level-2.
Rate limiting
A hard cap on write frequency prevents any failure mode — including ours — from oscillating a mill.
Watchdog and fallback
A 250 ms watchdog releases control deterministically to your existing Level-2 setup. The mill never stops because Steelira did.
Approval thresholds
Any change above your configured magnitude requires named human approval, recorded against the coil.
Immutable audit
Append-only, hash-chained decision log per coil, exportable and independently verifiable.
Data
Your grades stay your grades
Recipes and property windows are the most valuable IP in a steel plant. We treat them that way.
Tenant isolation
Every producer's data lives in a dedicated logical tenant with its own encryption keys, its own vector store namespace and its own model artefacts. There is no shared training pool and no cross-tenant retrieval path.
- Per-tenant keysEnvelope encryption with customer-managed keys available on Enterprise.
- No cross-trainingYour coils never train another producer's models. This is contractual, not aspirational.
exit gauge1.982 mm
flatness I-unit3.1 IU
F5 roll force18.4 MN
interstand tension9.6 MPa
chatter margin0.24 · safe
anneal energy-11.4%
pass 6/7autonomy: assisted-writetwin delta 0.004 mmaudit id 7f31c
Retention and residency
You choose the region and the retention period. Coil records, images and audit logs can be pinned to EU, US, India or Türkiye, or kept entirely on-prem for sensitive grades such as electrical steels and defence plate.
- Residency fixed per tenantEnforced at the storage layer, verifiable in the control plane.
- Exit on demandFull export in open formats at any time, contractually guaranteed.
residency eu-central (frankfurt)
retention_coils 7 years
retention_images 24 months
retention_audit 10 years (immutable)
cross_tenant disabled
export on-demand, json + parquet
Identity and access
SSO/SAML and SCIM provisioning with role separation between operator, rolling engineer, quality, security and admin. Approval rights and envelope-editing rights are separate roles by default.
- Least privilegeNobody can both widen an envelope and approve a write in the same role.
- Full attributionEvery approval carries a named identity, not a shared pulpit login.
exit gauge1.982 mm
flatness I-unit3.1 IU
F5 roll force18.4 MN
interstand tension9.6 MPa
chatter margin0.24 · safe
anneal energy-11.4%
pass 6/7autonomy: assisted-writetwin delta 0.004 mmaudit id 7f31c
Compliance
Certifications and frameworks
Status as of this release. We publish honestly, including what is still in progress.
| Framework | Status | Evidence available |
|---|---|---|
| SOC 2 Type II | [ASPIRATIONAL] In progress — Type I complete | Report under NDA |
| ISO 27001 | [ASPIRATIONAL] Implementation underway | Control matrix |
| IEC 62443 | Architecture aligned to SL-2 for the edge zone | Architecture review pack |
| GDPR | Compliant; DPA available | DPA, sub-processor list |
| Penetration testing | Annual third-party test | Executive summary under NDA |
| Vulnerability management | Continuous scanning, 30-day critical SLA | Policy and metrics |
[UNKNOWN] Certificate numbers and audit dates are withheld here and provided directly to prospective customers under NDA.
Safety
What happens when something goes wrong
Four failure modes, four deterministic answers.
- STEP 01
Model degrades
Confidence drops below threshold; the agent falls back from auto to assist and raises an alert. No silent degradation.
- STEP 02
Sensor drops out
Fusion continues on remaining signals with a widened uncertainty band; loops requiring the lost signal revert to Level-2.
- STEP 03
Edge node fails
Watchdog timeout at 250 ms releases control to your existing setup. The line keeps running as it did before Steelira.
- STEP 04
WAN is lost
The edge continues to perceive, decide and act locally, buffering telemetry. Nothing in the control path depends on the cloud.
AI assurance
Governing the models themselves
Model risk management, applied to a mill.
Every change is gated in CI
Model weights, prompts, envelopes and objective functions are version-controlled. Changes run against a golden dataset of your labelled coils and must clear accuracy and safety thresholds before they can be promoted to any line.
Grounding checks
Uncited claims in generated explanations are blocked, not softened.
Guardrails
Schema and safety validation on every agent output before it reaches arbitration.
Human checkpoints
Configurable per loop and per autonomy level, with named attribution.
Continuous evaluation
Golden datasets plus LLM-as-judge scoring on reasoning quality, sampled and human-reviewed by our metallurgists and yours.
Practices
How we run the company
Security is an engineering practice before it is a certificate.
Secure development
Mandatory review, dependency scanning, signed builds and reproducible edge images. No direct-to-production changes, ever.
Least-privilege operations
Production access is time-bound, approved and recorded. No standing credentials for customer environments.
Incident response
Documented severity model, 24-hour customer notification commitment for confirmed incidents affecting your data or control path.
Vendor management
Every sub-processor reviewed and published. Customers are notified before any sub-processor change.
Business continuity
Control-plane recovery objectives tested quarterly; edge autonomy means a control-plane outage does not stop a mill.
Responsible disclosure
security@steelira.com with a published policy and a commitment to acknowledge within one business day.
Proof
What OT teams say after review
“We ran Steelira in shadow for eight weeks on the tandem mill. It called eleven of the twelve off-gauge tail events before our AGC reacted. That was the moment the argument ended.”
“Our best roller retires in two years. Steelira is the first system that captured how he sets up a hard automotive grade — and then explained why, with the coil history to back it.”
“Prime yield moved 1.7 points on the galvanizing line and reheat gas fell double digits. Nothing else we bought this decade did both.”
Security FAQ
What security reviews always ask
The architecture is designed so that it cannot. Writes are clamped and rate-limited, a 250 ms watchdog releases control to your existing Level-2 on any fault, and the edge node has no inbound network path. The worst realistic failure is that Steelira stops helping.
We train models for your tenant on your data. We never train another producer's models on it, and that is a contractual commitment, not a policy statement.
Yes. Fully on-prem and air-gapped deployments are supported, with model updates delivered by signed media transfer.
Only the roles you designate — typically a rolling lead, a quality lead and an OT lead jointly. Promotions are recorded immutably with all signatures.
Request the security package
Architecture diagrams, OT segmentation detail, control matrix, penetration-test summary and DPA — sent under NDA within one business day.