Engineering write-ups, metallurgy arguments and commercial honesty from the team building rolling autonomy. Written by the people doing the work, not a content agency.
Read by rolling engineers at these producers
The most recent nine. Everything is public and ungated.
Automatic gauge control is a magnificent piece of engineering solving the wrong half of the problem. It corrects deviation it has already measured. Here is what changes when you predict it from entry profile and roll thermal state instead.
Every plant we visit has one or two people who can set up hard grades. Nothing in the plant captures how. We think the answer is not documentation — it is capturing disagreement between an engineer and a model, every day, for a year.
Strip-surface inspection at line speed is a systems problem long before it is a model problem. Frame ingestion, batching, TensorRT engine selection, and the three places latency hides that nobody profiles.
Every dispute we have had with a customer traced back to a vague baseline. Here is the methodology we now insist on before a single sensor is read, and why we let customer finance teams write half of it.
Mill chatter announces itself before it destroys a coil, in a signature that is measurable and grade-dependent. What the precursors look like, and why margin is a better control variable than a threshold.
Rare-event models need rare events. Using Omniverse Replicator and Cosmos to generate edge-wave, chatter and cobble precursors by the million, and how we validate that synthetic data is not lying to us.
Automatic gauge control is a magnificent piece of engineering solving the wrong half of the problem.
AGC measures exit gauge, computes an error and corrects it. That works beautifully in steady state and poorly at exactly the moments where prime yield is lost — threading, acceleration, tail-out and grade change — because in all four the deviation is created faster than a reactive loop can close it.
A twin approaches it differently. Given entry profile, roll thermal state, lubrication behaviour and the grade's deformation resistance, the deviation is largely predictable before the strip reaches the stand. The control problem becomes anticipation rather than correction, and the AGC gets a much easier job: holding a setpoint that is already close to right.
The measurable consequence on a design-partner tandem mill was a 38% reduction in off-gauge metres at head and tail, and gauge standard deviation on the coil body falling from 2.9 µm to 1.8 µm.
pass 6/7autonomy: assisted-writetwin delta 0.004 mmaudit id 7f31c
Reheat and anneal energy dominate a mill's footprint. A point of prime yield is a tonne of steel that never has to be made twice. The arithmetic linking yield, energy and CO₂ is more favourable than most plants have calculated.
The write-back contract, in full: signing, clamping, rate limiting, approval thresholds, the 250 ms watchdog and deterministic release. Published because your OT team should read it before they meet us.
Chemistry, slab, reheat profile, every pass, every image, every setpoint, every override and the final disposition — joined on coil ID. Most plants can assemble this in three days. It should take three seconds.
Five recurring threads.
Edge inference, control loops, connector archaeology and the latency nobody profiles.
TMCP paths, property windows, surface defect families and what a grade spec actually constrains.
Twin calibration, synthetic rare events, evaluation methodology and where our models are still wrong.
OT architecture, control-path guards and the write-back contract, published in full.
Baselines, outcome pricing and why we walk away from deployments that will not pay back.
Decarbonisation, the retiring roller, and where physical AI in heavy industry goes next.
Errors are corrected in place with a dated note. If we get something wrong about your mill, tell us and we will fix it publicly.
Three rules, applied to everything on this page.
We do not publish a number we cannot defend. Every performance claim on this site and in these posts traces to a design-partner engagement with an agreed baseline, and anything aspirational or unverified is marked as such.
We name our failure modes. The tag-mapping error that made our first shadow results look better than they were is in a published case study, because a vendor who only publishes wins is a vendor you cannot calibrate.
We do not write about competitors' products. Incumbents in mill automation, gauge control and surface inspection are excellent at what they do. Our argument is about scope and closed loops, not about their engineering.
“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.”
Most posts eventually come back to the connector layer, because most problems do.
Steelira writes to mill equipment, so security and safety are product requirements, not paperwork.
One email a month: what we published, what we learned on a line, and one thing we got wrong. No gating, no drip sequence, unsubscribe in one click.