Perception layer
Synchronised line-scan and RGB cameras, thermal imaging and laser gauges at the press exit, fused with furnace, press, quench, stretcher, saw and aging telemetry.
Platform
Perception, control and proofon one plant-edge runtime.
Extruon runs where the metal is. Edge agents fuse vision, thermal and laser gauging with press PLC telemetry, decide inside approved envelopes, write back to the line, and stream evidence to a cloud training loop that improves the next push.
Runs on plant-edge hardware · Degrades safely · Never depends on a live cloud link to keep the press safe.
Four layers
What the platformis actually made of.
No black box. Four layers, each independently inspectable, each with a defined failure mode.
Reasoning layer
Agents plan coupled moves across heating, press, die, quench, stretch, cut and aging under a plant orchestrator that arbitrates conflicting objectives.
Action layer
Bounded write-back into the press PLC, furnace and quench with envelopes, rollback and alarm interlocks tied to the existing fail-safe stop.
Assurance layer
Grounding and citation checks, human-in-the-loop checkpoints, continuous evaluation in CI, and an immutable action log for every decision.
Edge inference
Sub-second, at the press,whatever the network does.
The control loop fuses four to eight synchronised RGB or line-scan, thermal and laser-gauge streams per press while scoring defect, dimension, twist, temper-risk, recovery and energy models inside press-speed decision windows.
Target: sub-250 ms defect and dimension alerts, sub-1 second control recommendations at 30–120 FPS depending on camera type. [ASPIRATIONAL]
The runtime loop
Perceive. Plan. Push.Prove. Learn.
One closed loop runs at the plant edge on every billet. Nothing is advisory-only unless you want it to be — and nothing acts outside an envelope your engineers signed.
01Perceive
Sense the billet, the profile, the temper
Fused line-scan and RGB vision, thermal imaging, laser gauging and press PLC telemetry describe the push as it happens — dimensions, wall thickness, straightness, surface, exit temperature.
02Plan
Set the taper, the speed, the die, the quench
The agents plan billet-heating taper, ram speed, pressure and temperature, die-flow correction, quench rate, stretch, cut length and aging recipe as one coupled decision, not five isolated setpoints.
03Act
Run with adaptive control at the press
Approved moves write back into the press PLC, the furnace and the quench inside bounded action envelopes, with rollback and alarm interlocks wired to the same fail-safe stop your line already trusts.
04Prove
Predict defects and recovery before they cost metal
The loop predicts off-spec dimensions, surface defects, twist and temper misses, and flags die-trial risk early enough to change the push instead of scrapping it.
05Learn
Every correction trains the plant
Engineer approvals and corrections are logged to an immutable, assurance-grade audit trail and fed back into training — so the site's craft compounds instead of retiring.
The twin
Hit the profilebefore the push.
An Omniverse-based extrusion twin simulates billet heating, metal flow through the die, quench and resulting profile properties — then auto-optimises heating, press speed and die correction so a new die enters the press with corrections already applied.
- Simulation sweeps over billet temperature, ram speed, die temperature, quench rate and aging recipe.
- Physics-informed models calibrated against measured press outcomes, so the twin improves with every supervised push. [ASPIRATIONAL]
- Synthetic generation of rare die-line, blister, pickup, twist, bow and temper-miss variants — without waiting for expensive failures in production.
- Twin validation gates every increment of autonomy in CI before it reaches a live press.
# sweep before the first push
alloy 6063 temper T6
billet 178 mm length 1050 mm
taper 455 → 482 °C
ram_speed 5.8 → 6.9 mm/s
# predicted
exit_temp 506 °C in band
wall_dev ±0.05 mm in tolerance
twist 0.6 °/m stretch recoverable
recovery 88.4 %
die_trials 0 corrections pre-appliedAccelerated computing
Why this needsGPUs at the plant.
CPU-only deployment cannot hold the perception-to-action window. The stack is built on NVIDIA accelerated computing end to end.
Jetson / IGX
Plant-edge press-exit sensor fusion and robot handling inference, close to hot, fast lines where latency and network resilience matter.
Metropolis · DeepStream · TensorRT
High-frame-rate line-scan, RGB, thermal and laser-gauge pipelines for surface, profile, twist and temperature perception.
Holoscan
Real-time orchestration of thermal, laser, camera and PLC streams inside a bounded control loop.
Triton · NIM
Versioned, observable, fallback-capable multi-model serving: 8–20 models per plant with model-version routing. [ASPIRATIONAL]
Omniverse · Cosmos · Replicator
The press-and-die twin plus synthetic rare-fault generation — 50k–200k fault variants per product family. [ASPIRATIONAL]
cuOpt · RAPIDS
Press scheduling, billet and die assignment, cut nesting, rack loading, aging batches and energy windows; GPU ETL for fleet telemetry.
Autonomy ladder
Autonomy is earnedone gate at a time.
You choose how far up the ladder to go, and every rung has an exit. Nothing advances without measured accuracy, twin validation and your written approval.
Shadow mode
Read-only ingestion from press, furnace, gauge and vision, quench, saw, aging, MES and quality systems. Predictions are benchmarked against your operators and gauges with zero write access. [ASPIRATIONAL]
Assist mode
The agents recommend billet taper, ram-speed, exit-temperature, die-flow, quench, stretch, cut and aging moves. An engineer approves each one; the approvals train the models. [ASPIRATIONAL]
Bounded autonomy
Low-risk, customer-approved setpoint changes apply automatically inside action envelopes with rollback and alarm interlocks. Exceptions escalate to a human. [ASPIRATIONAL]
Scaled deployment
Expansion across presses, modules, sites and fleet-learning cohorts once ROI, safety and quality gates are met. [ASPIRATIONAL]
Architecture principles
Non-negotiableswe build against.
| Principle | What it means | What it guarantees |
|---|---|---|
| Multi-tenant isolation | Per-tenant data and model scoping; on-prem option for sensitive producers | Your die IP and quality history never mix with another extruder's |
| Event-driven agent runtime | Idempotent steps with human-approval checkpoints | A retried step never doubles a setpoint move |
| Model-agnostic router | Frontier and fine-tuned open models behind one interface | The best or cheapest model serves each step; COGS stays controlled |
| Security-first | SSO/RBAC, encryption in transit and at rest, SOC 2 programme | Passes IT, OT and security review before it touches the PLC |
| Scalable retrieval | Per-tenant vector stores with permission-aware filtering | Answers cite only sources the asker is allowed to see |
Integration
Connectors, nota rip and replace.
Extruon reads and writes through the systems you already run. If it speaks OPC UA, Modbus, MQTT or a REST API, we connect to it.
- Press PLC, billet-heating furnace and log shear, in-line profile gauge and vision, quench, stretcher, saw and aging oven.
- MES, quality and production-monitoring surfaces, plus SSO, ticketing and notification.
- Read-only in shadow mode; write-back enabled per setpoint, per press, per envelope.
# one press, one workflow, explicit envelopes
press:
id: "plant2-press3"
tonnage: 2750
sources:
- opcua://press3.plc/ram # speed, pressure, position
- opcua://furnace2/billet # taper zones 1-4
- rtsp://exit-linescan-01 # surface, 120 fps
- mqtt://gauge/profile # wall, width, twist
envelopes:
ram_speed: { min: 4.0, max: 8.5, max_step: 0.4 } # mm/s
exit_temp: { min: 495, max: 520 } # °C
taper_delta: { max: 30 } # °C front-to-back
autonomy: assist # shadow | assist | bounded
approver: "process-engineer"
rollback: on_alarmMeasured on the press
What the platformis measured on.
Every Extruon engagement starts with a baseline and ends with an audited delta. These are the target bands we underwrite in a paid pilot.
+4.2 pts
Press recovery uplift, saleable vs charged metal
−38%
Surface-defect and dimensional rejects
−61%
Die-trial pushes before a die is signed off
−12%
kWh per tonne across heating, quench and aging
Target outcome bands modelled from design-partner baselines. [ASPIRATIONAL — to be replaced with audited pilot results.]
Assurance-grade by default
Trust is the gateto autonomy.
Extruon never asks for control it has not earned. Shadow mode proves accuracy against your operators and gauges; advisory mode proves ROI; bounded autonomy only follows once the gates are met.
Per-tenant everything
Data, models, retrieval indexes and memory are scoped per tenant and per entity. On-prem deployment is available for sensitive producers.
Your die geometry stays yours
Die designs, customer drawings and recipes never leave your boundary. Fleet learning shares defect signatures and process priors, never geometry.
Immutable action log
Every perception, recommendation, approval and setpoint write is recorded with its evidence, ready for quality and customer audits.
SSO, RBAC, encryption
SSO and role-based access, encryption in transit and at rest, SOC 2 programme in progress, and fail-safe interlocks on every write path.
The finale
Put it on one press.Judge it on one number.
A paid pilot on your highest-pain workflow, with a defined recovery, scrap or energy success metric agreed before we start.
Land on one press. Expand press by press, module by module, site by site.