The questionsplants actuallyask.

A Physical AI platform that runs the foundry loop: factory-edge agents that sense melt, mold, pour, solidification and inspection state, plan setpoint moves, act inside approved process windows, verify results against inspection and retrain from every correction.

A system of action. Smelteon reads what your furnace, molding line, pouring cell, robots and inspection systems already produce, and — after shadow and assist modes and a safety review — writes bounded setpoint moves back into them.

Pre-revenue, recruiting three to five paid design partners around one measurable wedge workflow. Forward-looking financial, market and performance figures across this site are explicitly marked as aspirational.

Cell inference runs on Jetson-class hardware beside your PLC and robot controllers. Triton serving, the model router and RAPIDS ETL run on a factory server. Cloud is used for training, validation and fleet updates — never as a dependency for control.

Induction and holding furnaces, spectrometry and thermal analysis, molding and coremaking lines, auto-pour and ladle systems, HPDC and LPDC cells, KUKA, FANUC and ABB robot controllers, X-ray, CT, vision and dimensional inspection, foundry MES and ERP, and historians.

The cell keeps running. Local inference and bounded control continue, model promotion pauses, and the audit log buffers locally and replays on reconnect.

Nine: melt and chemistry, mold and core, pour and fill, solidify and cool, defect and inspect, yield and takt, finish and fettle, robot and handling, and quality and conformance — under a factory orchestrator, with the Castwin twin alongside.

Three gates. Every plan is checked against a process window your engineers authored and signed; every applied action carries an approver identity; and a last-known-good rollback path is armed for every controlled parameter, with automatic revert on outcome drift.

No. Plant safety systems and E-stops remain authoritative and independent. Smelteon never sits between an E-stop and a machine, and robot motion is bounded by existing safety-rated infrastructure as well as by our windows.

An immutable, append-only log of every plan, window check, approval, action, override, escalation, rollback and grading, tied to casting genealogy and designed as IATF 16949, ISO 9001, AS9100 and NADCAP evidence.

No. Recipes, geometries and programs are tenant-isolated and never leave your boundary. Federated fleet learning shares statistical priors about defects and process windows, never the recipes or geometry behind them.

Cell is $15,000 per melt, mold or pour cell per month. Factory is $90,000 per month for the whole-plant loop including the Castwin twin. Enterprise agreements are custom and land between $650k and $8.5M ACV.

The target is under 6–12 months on a cell where scrap, rework, melt-energy loss, inspection bottlenecks or labour constraints cost $500k–$5M a year, and under 14 months blended at enterprise scale.

No. A pilot is a paid engagement on the Cell tier against a success metric agreed in advance. We would rather charge for a pilot that proves a number than run a free trial that proves nothing.

A typical shadow-mode baseline runs about eight weeks, after which the golden-set accuracy review determines whether the cell can move to assist mode.