Multi-stream fusion
Video, thermal, acoustic, pressure-decay, torque-tool, PLC and functional-test streams fused per unit.
Everything theplatform does,in one place.
Perception, reasoning, action and assurance — plus the unglamorous features that decide whether a plant lets software write to a press: audit trails, rollbacks, isolation and evidence.
See every unit, not a sample
Statistical spot inspection was a compromise forced by human eyes and expensive cameras. Neither constraint applies any more.
Takt-speed inference
Sub-100 ms station decisions where the unit must be rejected or handled before it advances. Aspirational
Serialised unit record
Every unit carries its own perception, process and test history from blank to pallet.
Rare-fault coverage
Synthetic dent, seam, paint, wiring, foam, leak and robot-cell scenarios cover failures too rare or dangerous to collect.
Drift detection
Model and process drift surfaced against golden datasets before it becomes a yield event.
Assembly completeness
Presence, orientation and correctness of components, harnesses and electronics verified on every unit.
What adaptive control actually changes
Forming parameters
Press behaviour adapts to the coil, the tool and the product on the line rather than holding a fixed recipe until scrap appears.
- Blank-to-part parameter adaptation
- Seam and weld quality feedback
- Tool-wear compensation
Enamel and paint
Coverage, film build and cure adjust to maintain first-quality rate instead of chasing downgrades after the fact.
- Film-build targeting
- Run, sag and orange-peel prevention
- Cure-profile control
Foaming
Shot weight, mix ratio, jig conditions and cure time move to hold density and eliminate voids per cabinet.
- Density-distribution targeting
- Void-risk avoidance
- Thermal-performance forecast
Assembly and wiring
Torque, placement and connector verification gate the unit before it can advance, with evidence retained.
- Torque-tool closed loop
- Connector seating and polarity
- Rework routed with a cause
Test and triage
Leak and functional tests are sequenced, interpreted and triaged automatically, and failures are attributed upstream.
- Test sequencing via cuOpt
- Micro-leak signature detection
- Probable-cause rework routing
Line and robots
Line balance, changeover sequencing, rework routing and robot task allocation are optimised continuously.
- Takt protection
- Changeover sequencing
- Cobot allocation and zone safety
Grounded, cited, and explainable to an engineer
A recommendation a process engineer cannot interrogate is a recommendation they will not apply.
RAG over your reality
Retrieval over product and BOM specs, forming, foaming and assembly recipes, IEC 60335, energy and F-gas rules, and test procedures.
Model-agnostic router
The best or cheapest model serves each step; high-volume steps move to fine-tuned open models to control COGS.
Citation enforcement
Every output is traceable to a source document, recipe or measurement. No citation, no action.
Physics-informed models
Forming, paint, foam-density, leak, thermal and energy behaviour combined with learned models so advice stays explainable.
Per-engineer memory
Per-factory yield and defect history plus per-engineer and technician performance memory, scoped per tenant.
Guardrails
Schema and safety guardrails on every agent action, with continuous evaluation gating model and prompt changes in CI.
The seven agents in detail
Each is a product; together they are the factory.
Change the line in simulation before you change it in steel
The Omniverse-based line-and-product twin simulates forming, foaming, assembly and functional performance, and produces a predicted quality, yield and takt for a change before anyone touches the plant.
It is also how new product variants stop being a ramp-up scrap event, and how rare faults get into the training set without anyone having to cause one.
Twin capabilities
- Line layout, robot cell and foaming chamber simulation
- Product-variant validation before production release
- Changeover rehearsal and sequencing
- Cosmos / Replicator synthetic rare-fault generation
- cuOpt optimisation against your real constraints
- Predicted first-pass yield and takt as an acceptance gate
What Duromex talks to
Read and write, through the protocols your plant already runs.
| System | Direction | What Duromex does |
|---|---|---|
| Forming and stamping presses | Read + write | Adaptive parameters, seam quality feedback, tool-wear compensation |
| Paint and enamel lines | Read + write | Coverage and film-build control, defect classification, cure profiles |
| Foaming machines and jigs | Read + write | Shot weight, mix ratio, cure time, void and density prediction |
| Assembly stations and torque tools | Read + write | Placement, torque and connector verification with gating |
| Leak, functional and energy testers | Read + write | Test sequencing, micro-leak detection, automated failure triage |
| Appliance MES and quality systems | Read + write | Unit genealogy, dispositions, non-conformance and traceability |
| Identity (SSO/SAML/OIDC) | Read | Authentication, RBAC and per-tenant scoping |
| Ticketing and notifications | Write | Exception escalation to the people who own the line |
The features that get you write access
This is the half of the product nobody demos and every approver asks about.
Immutable audit log
Every agent action — what it saw, why it acted, what it wrote — recorded in an assurance-grade quality and safety trail.
Per-tenant isolation
Data, models, vector stores and memory scoped per tenant, with strict recipe and product IP protection.
Idempotent, reversible steps
Agent steps are idempotent with graceful degradation and fail-safe press, foamer, robot and line stop.
Evaluation gating
Golden datasets and LLM-as-judge evaluation gate every model and prompt change in CI before it can reach a line.
On-prem option
Sensitive producers can run entirely on premises, with cross-site learning federated rather than centralised.
Compliance evidence
IEC 60335 safety, energy-efficiency and refrigerant/F-gas evidence captured as a by-product of running the line.
Features are only useful in sequence
Perception without control is a report. Control without assurance is unapprovable. Assurance without a twin is slow. The features exist to make one sequence work, from blank to pallet.
Form & stampWeld & seamEnamel & paintFoam & insulateWire & assembleLeak & function testRate & pack
Engineering targets
The numbers the platform is built and measured against.
99.9%
uptime target
<100 ms
station decision latency Aspirational
60
sensor streams per line Aspirational
78%
target gross margin at scale
Which features actually get used
“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.”
Capability questions
Do you support our existing cameras and sensors?
Where they expose usable streams, yes — ingestion is designed to fuse heterogeneous sources. Where a station is under-instrumented, we will tell you what has to be added before a pilot can succeed.
Can the models be exported or run without Duromex?
Model artefacts trained on your data remain contractually yours in enterprise agreements. The runtime, orchestration and assurance layer are the product.
How do you keep inference cost under control?
The model router moves high-volume steps to fine-tuned open models, with caching and distillation over time. That is also how the 78% gross-margin target is reached.
Features are the answer. Yield is the question.
Tell us the number you are judged on. We will tell you which features touch it.
Duromex is pre-launch. Figures shown are design-partner targets and modelled economics, not audited results. Ask us for the methodology.