Perception engineer (vision)
Build line-scan, RGB and thermal pipelines on DeepStream and TensorRT that classify die lines, pickup, scoring and blisters at 120 FPS on a hot runout.
Careers
Write softwarethat moves 2,750 tonnes.
Extruon is a small team building physical AI for a process that has resisted software for fifty years. If you want your code to change what comes out of a die at 500 °C, this is the place.
Roles are hybrid with real time on plant floors. You will wear boots. [ASPIRATIONAL — hiring plan in progress.]
Open roles
Six roleswe are hiring for.
Control engineer
Design bounded control policies for ram speed, taper and exit temperature that a plant safety engineer will sign off — and that survive contact with a real press.
Edge runtime engineer
Rust and C++ on K3s and IGX: deterministic scheduling, sensor synchronisation, OTA with safe rollback, and graceful degradation when a camera dies mid-push.
Simulation engineer
Build the press-and-die twin in Omniverse: billet heating, metal flow, quench and temper, calibrated against measured outcomes until it can be trusted before a push.
Extrusion metallurgist
Own alloy, temper and die-design reasoning inside the product. Translate EN 755 and ASTM B221, die books and forty years of plant craft into models and evaluations.
Forward-deployed engineer
Live at design-partner plants. Install sensing, map tags, define envelopes, run baselines and turn engineer corrections into product requirements.
How we work
Six thingsabout working here.
Plant time is not optional
Everyone spends time on a plant floor. You cannot design bounded control for a press you have never stood next to.
Safety is a design constraint
Envelopes, interlocks and audit are not compliance chores. They are the features that make autonomy sellable.
Small team, deep scope
Perception, control, simulation and metallurgy in one room. You will work outside your specialty, on purpose.
Evidence over opinion
Baselines, golden datasets and twin validation settle arguments. Seniority does not.
Ship to a press, not to a demo
The definition of done is a shift running on your code with an audit trail behind it.
Customers are engineers
You will get feedback from press operators and die engineers directly. It is blunt and it is usually right.
The work
This is the loopyou would be building.
Perception at the exit, control at the ram, simulation before the push, and an audit trail behind every decision. Everything else is in service of it.
If this diagram makes you want to know what happens at breakthrough pressure, we should talk.
Interview process
Four steps.No brainteasers.
We test the work you would actually do, and we tell you what each stage is for before you do it.
- Intro call — what you have built, what you want to build, and whether the physical-AI problem excites you.
- Technical deep dive — a real problem from our domain: a perception pipeline, a control envelope, or a metallurgy trade-off.
- Plant context session — walk through a real push with our team and reason about what went wrong and why.
- Team and offer — meet the people you would work with, then a decision within a week.
- Equity
Meaningful ownership at an early stage. [PLACEHOLDER — bands published at offer.] - Hardware
Real GPUs, real sensors, real presses to test on. - Learning
Direct access to extrusion metallurgists and veteran press operators. - Scope
Own a module end to end, from sensor to setpoint to audit record.
What you will use
The stackyou would work in.
Rust · C++ · K3s · IGX
Deterministic plant-edge runtime with OTA and rollback.
PyTorch · TensorRT · Triton
Fine-tuned vision and time-series models served with version routing.
Omniverse · Cosmos · cuOpt
The twin, synthetic rare faults and constrained plant optimisation.
Go · Python · TypeScript
FastAPI services, a NestJS control plane and a Next.js console.
From the team
Why peoplejoin this problem.
I spent six years optimising ad ranking. Now a model I trained decides how fast a ram moves. The feedback loop is metal you can hold.
The hardest part is not the ML. It is earning the right to write a setpoint. That constraint makes the engineering better.
I came from thirty years in extrusion. Watching the craft get encoded instead of lost is the reason I took the job.
Design-partner quotes are illustrative of the target outcome. [PLACEHOLDER — to be replaced with named references at GA.]
Careers FAQ
Questions candidatesask us.
No — half the team came from software and learned the process. But you do need genuine curiosity about the physics, and a willingness to spend real time on plant floors.
Hybrid. Deep work happens anywhere; deployment, sensor installation and baseline work happen at customer plants. Forward-deployed roles travel substantially.
Small and deliberately so. You will own a module end to end rather than a slice of one. [PLACEHOLDER — current headcount on request.]
Earning write access. Making a model accurate is a known challenge; making a press engineer trust it enough to hand over the ram is the actual product.
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.
Measured on the press
What we areall working toward.
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.]
Careers
Build the autonomythe plant floor deserves.
Send us what you have built and which of the six roles fits. We reply to everyone.
Land on one press. Expand press by press, module by module, site by site.