Heat-and-press agent
Controls billet temperature taper, ram speed, breakthrough pressure, container and die temperature, and exit temperature to keep metal flow balanced and the profile on dimension.
AI agents
Agents that move metal,not just messages.
Edge and cloud agents run under a plant orchestrator with connectors into the billet-heating furnace, press PLC, in-line profile gauge and vision, quench, stretcher, saw, aging oven and MES. Each perceives a physical stage, plans a coupled move and acts inside an approved envelope.
Seven agents, one orchestrator
The plant runs as a crewof specialised agents.
Each agent owns a physical stage of the line, senses it directly, and acts inside an approved envelope. A plant orchestrator arbitrates between them so nobody optimises recovery at the expense of temper.
Profile-and-defect agent
Senses and predicts cross-section dimensions, wall thickness, straightness and surface defects — die lines, pickup, scoring, blisters — from fused vision and laser gauge.
Die-and-flow agent
Optimises die temperature, correction and flow balance to cut die trials and the twist and bow that send profiles back to the stretcher.
Quench-and-stretch agent
Controls press-quench cooling rate and stretch-straightening so temper and flatness land inside spec on the first pass.
Cut-and-age agent
Controls cut-to-length, batching and aging ovens to reach T5 and T6 mechanical properties without over-soaking the oven.
Robot-and-handling agent
Drives robotic pulling, stacking, racking and packing of long, hot, delicate profiles — the handling nobody wants to staff at 3am.
Yield-and-throughput agent
Optimises press recovery, scrap, butt and offcut loss, and energy per tonne across the whole line rather than one station at a time.
Extrusion-knowledge agent
Answers metallurgy and die-design questions with citations into your die books, profile drawings, press recipes and alloy specs.
Press-and-die digital twin
Simulates billet heating, metal flow through the die, quench and profile properties — hitting the target profile and yield before the push, not after the scrap.
Agent 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.
Orchestration
Somebody has toarbitrate.
Left alone, a yield agent will push speed until the surface degrades, and a quality agent will slow the press until the schedule dies. The plant orchestrator holds the objective function that makes the trade explicitly, with your weights.
- One objective function per plant: recovery, quality, throughput, energy and temper, weighted by you.
- Conflicts between agents are resolved before any setpoint is proposed, not after.
- Constraints from the twin and the envelopes are hard; preferences are soft.
- The arbitration is visible in the console — you can see which agent asked for what, and why it lost.
# proposals this push
yield-and-throughput ram_speed +0.6 mm/s recovery +0.4 pt
profile-and-defect ram_speed -0.2 mm/s die-line risk 0.31
quench-and-stretch hold quench rate T6 margin 0.8
# arbitration (weights: recovery .35 quality .40 energy .15 tput .10)
applied ram_speed +0.3 mm/s # inside envelope, step < 0.4
rationale die-line risk stays below 0.20 threshold
approver process-engineer · auto (bounded)
logged audit://plant2/press3/41208Model strategy
The right modelfor each step.
Models sit behind a router so the best or 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 vision models
Profile-dimension and surface-defect models — die lines, pickup, blisters — trained on your alloys, dies and profile families at the plant edge.
Thermal & laser-gauge sensing
Fused thermal and laser models for exit temperature field, wall deviation and geometric truth on the runout.
Metal-flow & exit-temperature models
Physics-informed predictors for flow balance and thermal response, calibrated against measured press outcomes.
Quench, temper & recovery models
Predict T5/T6 outcomes, flatness and recovery from the push, the quench and the aging recipe.
Time-series prognostics
Press, die and oven health forecasting for maintenance and die-life decisions.
Frontier reasoning models
Anthropic and Gemini models for extrusion-metallurgy and die-design reasoning and cited Q&A over your own die books.
Agents on the push
Seven agents,one billet.
Heat and press act. Profile and defect watches. Die and flow explains. Quench, stretch, cut and age finish. Yield and throughput keeps score. Knowledge answers why.
Each agent sees the same fused perception and writes only into its own bounded envelope.
Memory & retrieval
The agents rememberyour plant specifically.
RAG over your die books
Retrieval spans die books, profile drawings, press recipes, alloy and temper specs (EN 755, ASTM B221) and quality and surface procedures — multi-tenant isolated and permission-aware, with enforced citations.
pgvector + time series
Profile and defect image retrieval and die and recipe lookup in pgvector, alongside a time-series store for heating, press, quench and aging telemetry.
Per-plant, per-engineer
Recovery and defect history per plant, and performance memory per operator and die engineer — versioned and scoped per tenant, so the craft compounds without leaking.
Edge inference, cloud training
Plant-edge inference runtime with a cloud training loop, model versioning and extrusion-telemetry ingestion. Nightly incremental training targets 1M–10M sensor frames per enterprise customer. [ASPIRATIONAL]
Rare faults on demand
Cosmos and Omniverse Replicator synthesise rare die-line, blister, twist and temper-fault variations so the models see failures you cannot afford to produce.
Signatures, not geometry
Similar alloys, dies and profile families share defect signatures and process priors without exposing die geometry, drawings, recipes or plant performance. [ASPIRATIONAL]
Agent safety
What stops an agentfrom doing something stupid.
Autonomy without brakes is a liability. Four independent mechanisms sit between an agent and a press.
Bounded action space
Min, max and max-step per setpoint, defined by your engineers. Proposals outside the envelope are rejected at the boundary.
Approval checkpoints
High-impact, high-force-press and temper decisions require a human approval, configurable per press and per setpoint.
Simulation gate
Model and policy changes must pass twin validation and golden-dataset evaluation in CI before promotion.
Fail-safe stop
Alarm interlocks revert the last move and hand control back, wired into the existing press, quench, saw and robot safety chain.
From shadow to autonomy
How an agentearns write access.
Shadow benchmarking
The agent predicts and recommends with zero write access. Predictions are scored against operator decisions and gauge measurements on the same pushes.
Assist with approval
Recommendations reach the console. Engineers approve or correct each one. Approval rate becomes the readiness signal.
Bounded write-back
Low-risk setpoints inside tight envelopes apply automatically on approved presses, with rollback and full logging. Exceptions escalate.
Envelope expansion
Envelopes widen and more setpoints become writable as measured accuracy and safety hold. Every widening is a signed decision, reversible at any time.
Measured on the press
What the agentsare graded 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.]
Agent FAQ
The questionsOT engineers ask first.
Only in bounded autonomy, only on presses you have approved, only for setpoints you have made writable, and only inside envelopes you defined — with rollback on any alarm and a full audit record. In shadow and assist modes there is no write path at all.
The plant-edge runtime keeps running. Inference, envelopes and interlocks are local. Cloud connectivity is for training, fleet analytics and model distribution, not for keeping the press safe.
The plant orchestrator holds a single weighted objective function and arbitrates all proposals before anything is applied. The arbitration and its rationale are visible in the console.
Yes. Every proposal carries its evidence — the perception, the model version, the predicted effect and the rationale — and every answer from the knowledge agent carries citations into your own documents.
They correct it, and the correction is the training signal. Rising approval rate over time is the metric that governs whether autonomy expands.
The finale
Let them watchbefore they act.
Run the agents in shadow mode on your presses and score them against your operators and gauges. Write access is a decision you make later, with data.
Land on one press. Expand press by press, module by module, site by site.