The questions everyone asks

Your Level-2 sets up the mill and your AGC holds a setpoint reactively. Steelira predicts what the coil will do before the pass, coordinates gauge, shape, surface, properties, chatter and energy against one objective function, and only then asks your Level-2 to move. It makes the systems you already own act together.

No. It reads from it. Surface inspection systems are very good at detecting defects and structurally unable to act on the mill. Steelira closes that loop, and where your cameras are sufficient it uses them rather than asking you to buy new ones.

Shadow mode produces scored predictions against your own coils within six to eight weeks. That is the first honest result. Measured production improvement follows assist and bounded autonomy, typically three to five months from installation.

$18,000 per mill or line per month on the Line plan, $120,000 per plant per month on Plant, and custom enterprise agreements typically landing between $900k and $10M ACV. Annual prepay carries an 18% discount.

The architecture is built so it cannot. Writes are clamped per setpoint, grade and line, rate-limited, and a 250 ms watchdog releases control deterministically to your existing Level-2. 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 contractual. Grade recipes and property windows are the most valuable IP in your plant and we treat them accordingly.

Usually not for the first workflow. Most lines already have gauges, shape-meters, pyrometers and surface cameras. The common gap is vibration for chatter work, which is typically one accelerometer per stand.

The one with the highest product of tonnage, coil value and known loss — often the tandem cold mill or the galvanizing line. The free mill assessment answers this with your own numbers rather than ours.

In shadow and assist it costs a rejected recommendation, which becomes training data. In autonomy the envelope bounds the error and the watchdog releases control. Every case is logged, reviewed and fed back into the golden dataset.

Yes. Fully on-prem and air-gapped deployments are supported, with training on your own DGX or HGX cluster and model updates by signed media transfer.

You export it. Coil records, audit logs and model metadata in open formats, contractually guaranteed at any time. We would rather lose a renewal cleanly than hold data hostage.

Minimum and maximum deltas for each controllable setpoint, tightened per grade family, plus an approval threshold above which a named human must confirm. Your engineers author it, review it like any mill change, and sign it on merge.

Primetals, SMS X-Pact, Danieli Automation, TMEIC, ABB Ability and Siemens-based systems are covered by native connectors. Anything exposing OPC UA is supported by the generic path, which in practice covers almost everything else.

Median 46 ms from sensor frame to decision across 8–32 sensors per line, on a Jetson AGX Orin or IGX class node. Surface defect classification runs at around 12 ms, gauge risk under 4 ms.

Nothing changes in the control path. The edge node perceives, decides and acts locally, buffering telemetry until connectivity returns. Cloud connectivity is required for training and fleet management, never for control.

On the highest-value workflows, yes — a share of measured prime yield, energy or scrap improvement. We insist on the baseline methodology above first, which is precisely why we are willing to sign.

Typically 8–12 weeks at the Line rate, creditable against year one if you convert. We do not run free pilots because they do not get the attention they need from either side.

A rolling or quality engineer as champion, an OT contact for the network and write path, and someone from finance for the baseline. Roughly two to four hours a week during shadow, less afterwards.

Then it says that. We have told plants the wedge would not pay back at their tonnage, and we would rather do that at week zero than at month nine with a failed deployment on both our records.