Adaptive gauge control
Model-predictive AGC that anticipates the tail rather than reacting to it, holding tolerance through acceleration and threading.
Platform
The mill-edge autonomy stack
Perception, simulation, optimisation and bounded actuation in one runtime that lives beside the mill — GPU-accelerated at the edge, trained and governed from the cloud.
Runs on-prem or air-gappedDeterministic Level-2 fallbackOne model registry per group
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
Deployed alongside the automation vendors you already run
NordStrip Steel
Kalyon Metals
Ferrata Group
Aurum Flat Rolled
Meridian Coil
Volkan Çelik
Architecture
Four layers, one contract
Steelira is deliberately boring where it touches your mill and aggressive where it touches your data.
Sensing layer
Existing gauges, shape-meters, pyrometers, surface-inspection cameras, accelerometers and drive telemetry are ingested through OPC UA, direct camera links and historian taps. Where a line is under-instrumented, we specify the minimum sensor addition that unlocks the workflow — usually one camera bank or one accelerometer per stand.
- No rip-and-replaceYour IMS gauge and ISRA cameras stay exactly where they are.
- Time-alignedEvery signal is resampled onto a coil-length axis so head, body and tail are comparable across passes.
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
Edge inference layer
A Jetson Orin/Thor or IGX node per line runs DeepStream, Holoscan and TensorRT models for surface defects, thermal anomalies, flatness signals and chatter precursors. Triton serves the gauge-risk, property-risk and coating models side by side so one node covers the whole line.
- Sub-50 msMedian frame-to-decision latency across 8–32 sensors per line.
- Offline-tolerantThe node keeps controlling and buffering when the WAN drops; nothing about control depends on the cloud.
surface_defect_vit_b READY TensorRT 12.4 ms
flatness_shape_net READY TensorRT 6.1 ms
gauge_risk_tcn READY TensorRT 3.8 ms
chatter_precursor_cnn READY TensorRT 2.2 ms
property_tmcp_gnn READY ONNX 18.7 ms
coating_weight_reg READY TensorRT 4.0 ms
Decision layer
The mill orchestrator arbitrates between agents — Roll & Gauge wants tension, Roll & Chatter wants stability, Yield & Energy wants furnace setpoint down. Conflicts resolve against an explicit objective function you own, inside a safe operating envelope your engineers sign.
- You set the objectiveWeight prime yield, energy, throughput and roll life; the orchestrator optimises the weighted sum.
- BoundedHard clamps per setpoint, per grade, per line — enforced in the runtime, not in a prompt.
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
Actuation & assurance layer
Setpoints are written back through your Level-2 with signed requests, rate limits and a watchdog. Every action is logged against the coil ID with the perception, the twin prediction, the human approval and the measured outcome.
- Graduated autonomyShadow → assist → auto on low-risk loops → auto on the mill, promoted one gate at a time.
- Exportable auditQuality can pull the full decision trail for any coil, in a format that survives a customer claim.
14:02:11.204 perceive surface=clean shape=3.1IU
14:02:11.226 twin predict exit_gauge 1.981mm
14:02:11.238 plan F5 gap -0.018mm tension +0.4MPa
14:02:11.241 approve auto (envelope ok, gate L2)
14:02:11.259 act write L2 setpoint ack=true
14:02:19.880 verify measured 1.982mm Δ 0.001mm
Runtime loop
Perceive, plan, act, verify — 40 to 60 times a second
The loop is identical on a hot-strip finishing train, a tandem cold mill and a galvanizing line. Only the models and the envelope change.
- STEP 01
Perceive
Slab, strip, surface, shape, temperature, vibration and drive telemetry are fused at the mill edge into a single coil state, 40–60 times a second.
- STEP 02
Plan
The twin simulates roll-bite mechanics, thermal evolution, flatness and microstructure to pick the schedule, roll and TMCP path that hits gauge, flatness and properties.
- STEP 03
Act
Setpoints are written back through your Level-2 within signed, clamped limits — assisted first, then autonomous on the loops you promote.
- STEP 04
Verify
Every pass is scored against the twin's prediction and the coil's measured result, logged immutably, and fed back as training signal from your engineers' corrections.
Capabilities
What the platform does out of the box
Every capability ships with a validation harness so you can prove it on your own coils before it touches a setpoint.
Flatness & shape
Closed-loop bending, shifting and cooling control against a predicted shape profile, not just the measured one.
Surface defect sensing
Defect family, severity, coil-length position and probable root cause — classified at line speed, not in a Monday review.
TMCP path control
Finishing and coiling temperature targets planned per chemistry so microstructure lands in the property window.
Chatter & roll health
Third-octave precursor detection and roll-wear prognostics that move the roll change before it becomes a cobble.
Energy optimisation
Reheat and anneal dispatch optimised against the schedule so the furnace does not pay for a plan that changed.
Data platform
The coil history becomes an asset
Every pass, image, correction and outcome is written to a per-tenant store that makes the next coil better — and stays yours.
One record per coil, end to end
Chemistry, slab, reheat profile, every pass, every image, every setpoint, every human correction and the final quality disposition — joined on coil ID and queryable in seconds instead of three days of historian archaeology.
pgvector retrieval
Defect images and schedules retrieved by similarity, so the mill can ask “when did we last see this?”
TimescaleDB
Roll force, gauge, tension and temperature at native rate, retained for the life of the contract.
Versioned memory
Per-mill and per-engineer performance memory, scoped per tenant with no cross-tenant leakage.
RAG over the documents that govern the pass
Grade and property specs, ASTM/EN standards, coating specs, rolling schedules and quality procedures — retrieved with citations, so every recommendation traces to a source your metallurgist recognises.
Integrations
Speaks the mill's language on day one
Pre-built connectors to the systems already running your line — read for perception, write for control, with the same audit trail on both.
Level-2 & control
- Primetals Level-2
- SMS X-Pact
- Danieli Automation
- TMEIC mill drives
- ABB Ability
- Siemens SIMATIC / PCS 7
Gauge, shape & surface
- IMS X-ray gauges
- Vollmer thickness
- Shape-meter rolls
- ISRA Vision surface
- Parsytec inspection
- Thermal line scanners
Plant systems
- Level-3 MES / QMS
- SAP PP-PI
- OSIsoft PI / AVEVA
- OPC UA & Classic
- Historians (IP.21, Canary)
- CMMS / EAM
Interfaces
Read the mill, write the mill — same contract
A single typed API for perception, recommendation and bounded actuation. Everything the console does, your systems can do.
steelira/control.py
from steelira import Mill, Envelope
mill = Mill("hsm-1", site="kalyon-iskenderun")
# 1. read the live coil state (fused, coil-length aligned)
state = mill.coil_state(coil_id="84-2213-A")
print(state.exit_gauge_mm, state.flatness_iu, state.surface.defects)
# 2. ask the twin for the best next pass inside your envelope
plan = mill.plan_pass(
coil_id="84-2213-A",
objective={"prime_yield": 0.5, "energy": 0.3, "rate": 0.2},
envelope=Envelope(max_gap_delta_mm=0.05, max_tension_mpa=12.0),
)
# 3. write back through Level-2 — signed, clamped, audited
if plan.confidence > 0.92:
ack = mill.apply(plan, mode="auto", approver="agent:roll-gauge")
else:
ack = mill.propose(plan, to="pulpit-console") # human confirms
print(ack.audit_id, ack.applied_at)Deployment
Three deployment shapes, one product
Pick the shape your OT and security teams will actually approve.
| Shape | Where models run | Where training happens | Typical buyer |
|---|---|---|---|
| Edge + cloud (default) | Mill edge node per line | Steelira cloud, per-tenant | Most producers |
| Edge + private cloud | Mill edge node per line | Your VPC (AWS/Azure) | Groups with cloud standards |
| Fully on-prem | Mill edge node per line | On-site DGX or HGX cluster | Defence, electrical steel, sensitive grades |
| Air-gapped | Mill edge node per line | Offline, media-transfer updates | Export-controlled sites |
Control-plane availability targets and control-latency SLAs are contractual on Plant and Enterprise tiers.
Security
OT-safe by construction
Steelira writes to mill equipment, so security and safety are product requirements, not paperwork.
SOC 2 Type IIISO 27001IEC 62443 alignedGDPRSSO / SAMLData residency
- Segmented OT pathEdge runtime sits in a Purdue Level-2/3 DMZ with signed, rate-limited write-back and hard setpoint clamps.
- Deterministic fallbacksEvery control loop degrades to the mill's existing Level-2 setup on watchdog timeout — the mill never stops because Steelira does.
- Grade-IP isolationRecipes, grade specs and coil histories are tenant-scoped, encrypted at rest with per-tenant keys, and never used to train another producer's models.
- Assurance-grade auditEvery perception, recommendation, approval and actuation is written to an immutable, exportable log keyed to coil ID.
See the platform on your line
We will map your Level-2, gauge, shape and inspection stack against the connector library in one 90-minute session, and tell you honestly which workflow will pay back first.