Melt & Chemistry
Holds furnace melting, alloy chemistry, temperature, degassing and inoculation or nodularization on target — heat after heat, not shift after shift.
Nine agents.One factoryorchestrator.
Edge and cloud agents under a factory orchestrator, with melt and furnace, molding and coremaking, pouring and die-casting, inspection, robot and foundry-MES connectors. Each one perceives, plans against the twin, acts inside a window and logs everything.
9Agents
1Orchestrator
0Unlogged actions
Built with foundries and die casters melting iron, steel, aluminium, magnesium and copper
Ironhaus GroupMeridian CastingsVulcan DuctileAurora Light MetalsKestrel Die CastingNorthfield Foundry Co.
[PLACEHOLDER] Design-partner names are illustrative until first references are signed.
What each agentowns.
An agent is not a chatbot over a historian. It is a bounded controller with its own perception, its own process window and its own approval gate.
Mold & Core
Controls sand molding, coremaking, sand properties and mold quality so geometry and strength arrive right the first time.
Pour & Fill
Drives robotic pouring: ladle and stream, pour temperature, timing and mold fill — the most craft-bound, most dangerous step in the plant.
Solidify & Cool
Controls solidification, cooling curves and shakeout timing, the window where shrinkage porosity and hot tears are decided.
Defect & Inspect
Senses and predicts gas and shrinkage porosity, inclusions, cold shuts, misruns and hot tears from X-ray, CT and vision.
Yield & Takt
Optimizes yield, scrap, melt loss and line balancing across melt, mold and pour so the line runs to takt.
Finish & Fettle
Runs robotic grinding, cut-off and deburring to spec, hands-free, and returns gate and riser data upstream.
Robot & Handling
Commands pouring robots, mold handling and casting handling inside bounded, safety-reviewed envelopes.
Quality & Conformance
Owns right-first-time, non-conformance, casting genealogy and traceability under IATF 16949, ISO 9001, AS9100 and NADCAP.
Perceive, plan, act,verify, log, retrain.
The same six-step loop runs at every station. The only thing that changes is what is being sensed and what is being moved.
-
Perceive
Melt, sand, mold, pour stream, thermal history and casting surfaces are read continuously at the cell — furnace spectrometry, sand lab, pour vision, thermal curves, X-ray, CT and dimensional scans.
-
Plan
Castwin simulates the pour and solidification for this alloy, geometry and mold state, then proposes melt chemistry, gating, pour and cooling setpoints that hit the soundness, yield and takt target.
-
Act
Bounded write-back moves furnace, molding, pouring and robot setpoints inside a process window that a metallurgist approved — never outside it.
-
Verify
Poreon grades the casting against the prediction: porosity, inclusions, cold shuts, misruns, hot tears, dimensions and microstructure.
-
Log
Every action, approval, override and outcome lands in an immutable, assurance-grade quality and safety audit log tied to the casting's genealogy.
-
Retrain
Corrections from metallurgists and inspection outcomes train the models — the melter's, molder's and pourer's craft, encoded and compounding.
RouterEdge modelsCloud modelsEvals
Models28
p95 edge41ms
Golden set94.6%
poreon-ct-v9 · TensorRT · edgeserving
meltra-chem-v4 · fine-tuned openserving
metallurgy-reasoning · frontieron demand
eval gate · CI blocked promotion v10held
The right modelfor each step,at the right cost.
Models sit behind a router so the best and cheapest model serves each step. High-volume steps move to fine-tuned open models to control COGS; premium reasoning steps use frontier models.
- Fine-tuned porosity, inclusion, cold-shut and misrun vision models at the edge
- X-ray and CT fusion models on Jetson and IGX
- Melt-chemistry, sand and solidification physics-informed models
- Time-series prognostics for furnace, molding and pouring equipment
- Frontier models (Anthropic, Gemini) for casting-metallurgy and safety reasoning and Q&A
The craft, encoded.
Retrieval and memory are what make an agent a foundry engineer rather than a generic assistant.
Vector store
pgvector for defect, porosity and X-ray/CT image retrieval plus recipe and geometry lookup, with a time-series store for melt, molding, pouring, solidification and yield telemetry.
Grounded RAG
Retrieval over melt and sand recipes, gating and solidification specs, casting standards (IATF 16949, ISO 9001, AS9100, NADCAP, ASTM) and quality procedures — multi-tenant isolated, permission-aware, citation-enforced.
Plant and person memory
Per-plant yield-and-quality history plus per-melter, pourer and engineer performance memory — the scarce craft, versioned and scoped per tenant so it improves without leaking.
An agent nearmolten metalearns every inch.
01Citations or it does not ship
Every output carries a grounding check. A recommendation that cannot cite the recipe, spec or standard it came from is escalated, not applied.
- Citation and grounding checks on every output
- Guardrails for safety and schema
- Unsourced answers escalate to a human
- Grounding
- Evaluation
- Authority
Ask an agentfor a plan.
Plans are objects, not prose: a window, a confidence, an explanation and an outcome that gets graded later.
# Ask the pour agent for a bounded setpoint move on cell P-04
from smelteon import Plant
plant = Plant("plant-04", edge="foundry-edge-02")
heat = plant.heats.current(line="ductile-iron-1")
plan = plant.agents.pour.plan(
heat = heat.id,
part = "crankshaft-4cyl-rev-c",
mold_state = plant.agents.mold.state(),
objective = "minimise shrinkage porosity at journal 3",
twin = "castwin:solidification-v7",
)
# Every move is checked against the approved process window first
if plan.within_window and plan.confidence > 0.94:
plant.apply(plan, mode="bounded", approver="metallurgist-on-shift")
else:
plant.escalate(plan, reason=plan.explain())Who can move what,and under whose approval.
| Agent | Senses | Can move (bounded) | Approver |
|---|---|---|---|
| Melt & Chemistry | Spectrometry, thermal analysis, furnace state | Alloy additions, power profile, tap temperature | Metallurgist |
| Mold & Core | Sand lab, compaction, core cell, mold scan | Sand additions, shot parameters, mold reject | Process engineer |
| Pour & Fill | Pour vision, ladle telemetry, mold state | Tilt profile, pour temperature, fill time | Metallurgist + safety |
| Solidify & Cool | Thermal curves, line position | Cooling line speed, shakeout timing | Process engineer |
| Defect & Inspect | X-ray, CT, vision, dimensional | Disposition routing, hold and release | Quality engineer |
| Yield & Takt | Line state, order book, furnace readiness | Schedule, mold allocation, energy windows | Operations |
| Finish & Fettle | Casting scan, gate and riser geometry | Robot paths, grind parameters | Process engineer + safety |
| Robot & Handling | Cell vision, safety envelope state | Motion inside safety-rated envelopes | Safety / EHS |
| Quality & Conformance | Genealogy, non-conformance, audit log | Records, holds, conformance evidence | Quality engineer |
9
Agents under one orchestrator
6
Steps in every agent loop
3
Autonomy gates before write-back
100%
Actions in the immutable audit log
What quality andsafety ask first.
Three things: it can only propose moves inside a process window a metallurgist approved, it requires an approver identity to apply anything in bounded mode, and every action is logged with a rollback path armed. Anything outside the window escalates instead of acting.
Model priors can be shared through federated learning; your alloy recipes, geometries, programs and telemetry are tenant-isolated and never leave your boundary. On-prem deployment is available where that is a requirement.
It becomes a labelled training example of safe control under real plant constraints. That corpus — approved and rejected actions alike — is the moat, and it is why the system gets better every heat.
Put one agenton one cell.
Shadow mode, your baseline, your inspection results as ground truth. Autonomy graduates only when the numbers say it should.