Trust before autonomy
Grounding, audit, human-in-the-loop and graduated control come first. We would rather deploy slowly than lose a mill's confidence once.
Nearly two billion tonnes a year become the cars, appliances, buildings and pipelines the modern world runs on. Almost all of it is shaped by rolling — and rolling is still a craft. We are building the layer that makes it a system.
pass 6/7autonomy: assisted-writetwin delta 0.004 mmaudit id 7f31c
Working with producers on four continents
Rolling is the last black art in a trillion-dollar industry.
Rolling is one of the most extreme thermomechanical processes on earth. A slab is reduced from centimetres to millimetres in seconds while roll force, inter-stand tension, temperature and metallurgy interact non-linearly at speed. Then it is cold-rolled to microns, annealed, tempered and coated. Every step decides whether a coil is prime steel or downgraded scrap.
The industry has automated the parts that are easy to model and left the rest to people. Level-2 systems hold setpoints. AGC reacts. Inspection systems flag defects nobody can act on in time. Simulation runs offline, hours after the coil has shipped. In between sits a master roller whose intuition is genuinely irreplaceable — and who is retiring.
Three things changed at once. GPU inference got fast enough to run vision and physics models at mill speed on a node beside the line. Simulation got good enough to predict flatness and microstructure before the pass, not after. And decarbonisation made energy and yield existential rather than merely important. That is why this company exists now rather than in 2015.
We are not building a dashboard. Mills have dashboards. We are building a system of action: something that perceives the coil, decides the pass, writes the setpoint and proves the result — with a human checkpoint wherever the plant wants one, and an audit trail that survives a customer claim.
Four commitments that constrain what we ship.
Grounding, audit, human-in-the-loop and graduated control come first. We would rather deploy slowly than lose a mill's confidence once.
We build for steel and flat metals, not for generic industry. The models, the connectors and the vocabulary are specific on purpose.
We are willing to price on measured prime yield, energy and scrap — which means we are careful about baselines.
Every correction a master roller makes is training signal. Capturing that craft before it retires is the moat and the mission.
Clamps, watchdogs and deterministic fallback are architecture. Safety is not a feature we can descope.
Where something is aspirational or unverified, we mark it. Heavy industry buyers can tell the difference anyway.
Deliberately unglamorous milestones.
Formed around a thesis: the rolling mill is the highest-value unsolved control problem in heavy industry.
Structured assessments across hot-strip, tandem cold and coating lines to find where the losses actually sit.
[ASPIRATIONAL] First lines instrumented in read-only mode with baselines agreed by customer finance teams.
[ASPIRATIONAL] Recommendations approved on the pulpit, with acceptance rate tracked toward promotion gates.
[ASPIRATIONAL] First loops promoted to autonomous control inside signed envelopes.
[ASPIRATIONAL] Full-plant deployments with the mill-and-metallurgy twin planning schedules across lines.
[UNKNOWN] Headcount, funding and legal-entity details are provided directly to counterparties rather than published here.
Three disciplines that rarely sit in one room, which is exactly the point.
People who have run a pulpit, set up a hard automotive grade and cleaned up after a cobble. They define what correct means and they hold veto over anything that reaches a mill.
Vision, time-series and physics-informed modelling, edge inference, and simulation. The people who make a decision land in under 50 milliseconds.
Segmentation, connectors, fleet management and the boring, essential work of making a system safe enough to write to a hot mill.
In the plant, every week, through shadow and assist. Most of what the product knows came from them arguing with a rolling engineer on a Friday.
Twin calibration, synthetic rare-event generation and evaluation methodology. They own the golden datasets and the gates.
Small, technical, and instructed to walk away from plants where the wedge does not pay back. A bad deployment costs more than a lost deal.
From one line to every coil.
The first version controls gauge and shape on one line. The second controls the whole plant — schedule, roll, inspect, anneal, temper, coat — against one objective function the plant director sets in the morning.
After that, the interesting part. Every mill running Steelira contributes supervised corrections to its own models, and the methodology for capturing that craft transfers between plants even when the data never does. Rolling becomes a discipline with a shared, improving body of encoded knowledge instead of a craft that resets with every retirement.
The long-term goal is straightforward to state and hard to earn: become the autonomous operations layer for the world's steel and metals mills, so that every coil is rolled, annealed, coated and shipped on Steelira.
“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.”
Steelira writes to mill equipment, so security and safety are product requirements, not paperwork.
We are hiring metallurgists, control engineers and physical-AI researchers who want their work to show up in a coil rather than a benchmark.