Edge and cloud agents run under a factory orchestrator with connectors into the stamping press, paint and enamel line, foaming machine, assembly stations, leak and functional testers and the appliance MES. They do the work. Humans supervise the exceptions.
Fused vision, thermal, acoustic, pressure-decay, torque-tool, PLC and functional-test streams describe the panel, the surface, the assembly and the leak path on every unit — not a sample.
The orchestrator plans the form-and-finish, foam, assemble-and-wire and leak/functional-test sequence for the product actually on the line — including model changeovers.
Adaptive control writes back into the press, the enamel line, the foaming machine, the assembly stations and the testers — within the authority the plant grants.
Surface, weld and assembly defects, foaming voids, refrigerant and water leaks and functional performance are predicted before the tester confirms them.
First-pass yield, scrap, rework, energy and line balance are optimised continuously across the factory, not tuned once per quarter.
Every action lands in an immutable, assurance-grade quality and safety audit log. Every engineer correction trains the models.
Seven agents, one per stage of the build, sequenced by one orchestrator against the product on the line.
An agent perceives, decides, acts, and is accountable for the result. Everything below is what that requires in a factory.
Decisions are made from fused, serialised sensor evidence attached to a specific unit, not from an aggregate.
The orchestrator plans a sequence against the product, BOM and recipe actually on the line — including the changeover.
Agents write into presses, foamers, stations and testers, within an authority envelope the plant grants per line.
High-impact, safety and warranty decisions stop for a human. Every correction becomes training signal.
Every action is logged immutably with its evidence and its citation chain. Nothing is anonymous.
Each agent is scored against the human and tester baseline it is meant to beat, continuously.
Form-and-Finish shapes and finishes the cabinet. Defect-and-Surface watches every panel. Foam-and-Insulate fills it. Assemble-and-Wire completes it. Leak-and-Test proves it. Robot-and-Handling packs it. Yield-and-Throughput makes the whole sequence pay.
Fine-tuned specialist models at the edge; frontier reasoning where judgement and explanation matter.
| Agent | Primary models | Where it runs |
|---|---|---|
| Form-and-Finish | Forming/press time-series + seam vision | Factory edge |
| Defect-and-Surface | Fine-tuned surface, weld, dent and enamel defect vision | Factory edge |
| Foam-and-Insulate | Foam density/void models + thermal performance regression | Factory edge |
| Assemble-and-Wire | Assembly-completeness vision + torque/connector verification | Factory edge |
| Leak-and-Test | Leak-signature fusion + functional/energy performance models | Factory edge |
| Robot-and-Handling | Isaac-based robotics policies + zone safety | Factory edge |
| Yield-and-Throughput | cuOpt optimisation + yield prognostics | Edge and cloud |
| Appliance-Knowledge | Anthropic and Gemini reasoning over RAG | Cloud (or on-prem) |
Per-factory yield and defect history plus per-engineer and test-technician performance memory capture the scarce craft of appliance process engineers — the knowledge that currently retires out of the building.
Memory is versioned and scoped per tenant and per entity, so it improves over time without any cross-tenant leakage.
Nothing is granted. Everything is measured.
Predict alongside the humans. Publish the accuracy gap against the measured baseline. Write nothing.
Recommend and predict; a human approves each action. Corrections train the model in the loop.
Act automatically on low-risk lines and test workflows. Escalate exceptions. Authority is revocable per line.
Any recipe, layout or changeover change is validated in the line-and-product twin before it reaches production.
Golden datasets and judge models gate every model and prompt change in CI, and drift is monitored in production.
Duromex depends on accelerated compute because decisions must land inside takt while rare failures are simulated and trained elsewhere. CPU-only infrastructure cannot carry this workload.
Station-level inspection for panel surface, weld, dent, enamel, connector, wiring, assembly completeness, robot-zone safety and pack checks, targeting sub-100 ms where takt demands it. Aspirational
Video, thermal, acoustic, pressure-decay, torque-tool, PLC and functional-test streams fused for leak, foam-void, assembly and energy-performance prediction. Aspirational
Vision, time-series, RAG and optimisation models served from factory-edge GPU clusters with versioning, canary rollout, rollback and per-tenant isolation. Aspirational
Defect, leak, foam-density, functional-performance, yield and reliability models retrained from supervised outcomes, engineer corrections and warranty feedback. Aspirational
Line layouts, robot cells, foaming chambers, test stations and new variants simulated, with 10,000+ synthetic rare-fault scenarios per product family. Aspirational
Every agent carries a number it must beat.
baseline per workflow, measured in shadow
high-impact actions with a human checkpoint
uncited outputs permitted
gates before production: CI evaluation and twin
“Our end-of-line functional testers told us what failed. They never told us which press stroke or foam shot caused it. That link is the whole product.”
“Foam voids are invisible until the energy rating comes back wrong. Predicting density distribution before the cabinet cures is the part I could not buy anywhere else.”
“If it writes to the press, it needs an audit trail a quality auditor accepts. Duromex started there instead of bolting it on.”
It escalates. Confidence thresholds are set per workflow and per risk level; below threshold the unit is routed to a human with the evidence and the probable cause attached.
Fail-safe stop paths exist for press, foamer, robot and line handling, and they are part of the safety design rather than an agent decision. Agents request; the safety layer decides.
They replace the parts of the job that are repetitive perception and reactive rework. The engineer becomes the supervisor whose corrections are the most valuable input the system receives.
Continuous evaluation against golden datasets, drift monitoring, twin validation before release, and versioned models that can be rolled back per tenant.
Shadow mode, one metric, and a published accuracy gap. That is the whole ask.
Duromex is pre-launch. Figures shown are design-partner targets and modelled economics, not audited results. Ask us for the methodology.