Roll & Gauge agent
Holds exit gauge to microns by trimming roll gap, roll force, bending, shifting and inter-stand tension pass by pass — including the head and tail ends where prime yield is lost.
AI agents
Seven agents. One mill orchestrator.
Each agent owns a real mill responsibility, holds its own models and memory, and negotiates with the others against an objective function you control. None of them can move a setpoint outside the envelope your engineers signed.
Bounded by an explicit envelopeEvery decision attributableBuy one, or buy all seven
roll-gauge propose F5 gap -0.018mm gain +0.6 prime
roll-chatter veto gap step > margin margin 0.19
roll-gauge revise F5 gap -0.011mm gain +0.4 prime
yield-energy concur furnace -6°C gain -0.8% gas
orchestrator commit weighted objective 0.5/0.3/0.2
audit decision 7f31c signed, envelope ok
Agents running on design-partner lines today
NordStrip Steel
Kalyon Metals
Ferrata Group
Aurum Flat Rolled
Meridian Coil
Volkan Çelik
The agents
What each agent owns
Agents are products. You can start with one.
Surface & Shape agent
Fuses high-speed strip-surface vision with shape-meter and thermal data to detect and predict scale, scratches, slivers, inclusions, roll marks, edge-wave and centre-buckle.
Schedule & Flow agent
Optimises the rolling schedule, reduction split and thermomechanical path so the coil hits target properties in fewer passes, with campaign and roll-change planning on cuOpt.
Temper & Coat agent
Controls annealing, temper/skin-pass, pickling and coating weight so mechanical properties and coating spec land inside the window without over-coating.
Roll & Chatter agent
Predicts roll wear, thermal crown and third-octave chatter precursors to schedule roll changes early and steer the mill away from instability and cobbles.
Robot & Handling agent
Drives robotic coil handling, strapping, sampling and lab logistics on Isaac-based cells so sample-to-result time stops gating the next campaign.
Yield & Energy agent
Optimises prime yield, scrap and cobble risk against reheat and anneal energy — the two levers that decide margin and CO₂ on the same coil.
Orchestration
Agents disagree. That is the point.
Roll & Gauge wants tension. Roll & Chatter wants stability. Yield & Energy wants the furnace six degrees cooler. A single model would average those away; an orchestrator makes the trade-off explicit, weighted by an objective you own and logged so anyone can audit why the mill did what it did.
- Explicit objective functionWeight prime yield, energy, throughput and roll life. Change the weights per campaign, per grade or per shift.
- Veto rightsSafety-critical agents can veto, not just vote. Chatter margin beats a yield gain, always.
- AttributionEvery committed decision records which agent proposed it, which objected, and on what evidence.
- Deterministic tie-breaksNo stochastic behaviour in the control path. The same state and objective produce the same decision.
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
Agent loop
How a single agent thinks
Identical structure across all seven — which is why adding the eighth is a model problem, not an architecture problem.
- 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.
Configuration
Agents are configured, not prompted
Envelopes, objectives and autonomy levels live in version-controlled configuration that your engineers review like any other mill change.
mills/hsm-1/agents.yaml
# Steelira agent configuration — reviewed via PR, signed on merge
mill: hsm-1
objective:
prime_yield: 0.5
energy: 0.3
throughput: 0.2
agents:
roll_gauge:
autonomy: auto # shadow | assist | auto
envelope:
gap_delta_mm: [-0.05, 0.05]
tension_mpa: [4.0, 12.0]
approval_required_above: 0.03
roll_chatter:
autonomy: auto
veto: true # can block any proposal
margin_floor: 0.20
temper_coat:
autonomy: assist # human confirms every move
grades: [DX51D, DX54D, S350GD]
fallback:
watchdog_ms: 250
on_timeout: "release_to_level2"Autonomy levels
Four levels, promoted one gate at a time
Promotion is a decision your team makes on evidence, not a switch we flip.
| Level | What the agent does | What the human does | Typical gate to promote |
|---|---|---|---|
| Shadow | Observes and predicts only | Nothing — production is unaffected | 6–8 weeks, accuracy vs. measured outcome |
| Assist | Proposes every move | Approves, adjusts or rejects each one | ≥90% acceptance rate over 4 weeks |
| Auto (bounded) | Acts within a tight envelope | Supervises; handles exceptions | Zero envelope breaches, ROI confirmed |
| Auto (mill) | Acts across coordinated loops | Sets objective; reviews campaigns | Signed by rolling, quality and OT leads |
Grounding
Why a metallurgist believes it
An agent that cannot show its reasoning will never be allowed near a hot mill.
Every recommendation cites its evidence
The coil histories it matched, the grade spec and standard it reasoned from, and the twin run that predicted the outcome — attached to the decision, not buried in a log.
Citations enforced
Grounding checks run on every generated output; uncited claims are blocked, not softened.
Golden datasets
Model and prompt changes are gated in CI against your labelled coils.
LLM-as-judge
Continuous evaluation of reasoning quality on metallurgy questions, sampled and human-reviewed.
The roller's craft, encoded
Per-mill and per-engineer memory captures how your best roller sets up a hard grade — versioned, tenant-scoped, and still there after he retires.
Proof
Working with the agents
“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.”
Agent FAQ
What people ask before they let an agent write
Not to the mill. Agents can only emit setpoint changes inside a per-setpoint, per-grade, per-line envelope enforced in the runtime, and any change above your configured threshold requires human approval. The control path is deterministic.
No. Control is handled by fine-tuned vision, time-series and physics-informed models with deterministic arbitration. Language models are used for metallurgy reasoning, operator Q&A and explanation — never in the control path.
Yes. The Line plan is one mill or line with one agent family, which is how almost every deployment starts.
In shadow and assist it costs nothing but a rejected recommendation, which becomes training data. In auto, the envelope bounds the error and the watchdog releases control to your Level-2. Every case is logged and reviewed.
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.
Put one agent in shadow mode
Start with the agent that owns your biggest loss. Read-only, six weeks, scored against your own coils.