Quality 4.0 · CNC production

Catch the defect before the machine makes it.

Undrd builds quality systems for factories: AI inspection that sees what rule-based vision misses, and closed-loop control that corrects the machine while the part is still in the chuck. Retrofitted onto the controllers you already own.

See what we build Talk to an engineer

In-process correction

Nominal

Ø40.000

Measured

Ø40.026

Tolerance

±0.030

Correction written

−0.013

radius, X wear

Measured on the machine. Corrected before the next part is cut.

Rule-based vs learned inspection

A rule catches the defect it was written for. A model catches the one nobody saw coming.

Most factory inspection still runs on thresholds somebody tuned by hand. That solves what it was built to solve: a known feature, in a known place, measured the same way on every part. What it cannot do is recognise a defect nobody wrote a rule for — and it only looks once the part is finished.

Industry 3.0

Rule-based, end of line

Hand-coded if/then thresholds, written once per defect type
Re-tuning every time the lighting, fixture or part orientation shifts
Blind to any defect nobody thought to write a rule for
Judges finished parts, so by the time it rejects one the material is already cut

Industry 4.0

Learned, in-process

One model learns the defect space from examples instead of from rules
Image and process signals judged together, so a part that looks clean but ran outside its process window is still caught
Unfamiliar failure modes get flagged rather than slipping through
Runs while the part is being made, so drift is corrected before it becomes scrap

Capabilities

What we build

Six things we do. Most plants start with one and add the others over time.

AI part inspection

A multi-modal model that judges the image and the process signature at the same time.

A vision branch reads the camera frame while a second branch reads temperature, pressure, vibration, cycle time and dimensional deviation — both fused into a single verdict
Catches the failure that single-signal systems miss: the part that photographs clean but ran outside its process window
Every prediction reaches the dashboard with its defect class and confidence
Defect frames are kept, so an operator can see exactly what the model saw

PyTorchImage + sensor fusionDefect classification

Closed-loop tool-wear correction

The part is measured, the offset is recalculated, and the machine is corrected — automatically, every cycle.

Reads the finished diameter straight from the measuring machine or an in-process probe
Computes the radius correction, then applies a deadband, a per-cycle clamp and a hard per-tool wear limit before anything is written
Writes the offset back as a G10 block over RS-232, or straight into a Modbus holding register
Every correction is logged per tool, so wear becomes a trend you can watch instead of a surprise

G10 L10Modbus TCPPer-tool wear limits

Legacy machine connectivity

Bring machines built before "Industry 4.0" was a phrase onto the network — without replacing them.

Serial, fieldbus or a vendor API — the first job is working out what your controller will actually expose, because that varies by model and by vintage
Every controller sits behind one fixed interface, so a machine we have not met before is a contained adapter rather than a rewrite
Reads measurements out and writes corrections back on the same link, with the protocol and timing matched to the machine
Where a controller offers no way in at all, a GPIO relay can drive external cycle start
The edge agent runs on a Raspberry Pi or Revolution Pi beside the machine and keeps working when the network doesn't

RS-232Modbus TCPVendor APIsIn-process probing

Safety and piece watchdogs

Vision watchdogs that stop the line for what a fixture sensor cannot see.

Person detection inside a configured zone signals STOP, and RESUME once the zone has been clear for a set number of frames
A three-class piece watchdog separates absent, correctly seated and incorrectly seated parts before the cycle starts
Hysteresis on every signal, so one bad frame never trips the line
Deployed alongside your existing interlocks, not in place of them

! This is a software watchdog. It complements certified safety hardware (ISO 13849 / IEC 61508) and must never be the only safety mechanism on a machine.

YOLOv8nPiece presencePLC stop / resume

Fleet dashboard and quality analytics

One view across every plant, scoped to what each person is allowed to see.

Organisations, plants and devices, with separate scopes for system administrators, senior managers, managers and operators
A rolling alert engine over a 30-minute window, with acknowledgement and delivery by email or in-app
Quality events with a formal disposition, and an audit log of who changed what
Throughput, pass rate and defect breakdown — per device, per plant, across the fleet
English and Portuguese throughout

FastAPIPostgreSQLRole-scoped access

NC programs and toolpath verification

Manage the programs your machines actually run, and check them before they cut.

A per-device program store with operation metadata, probe links and per-tool enable or disable
Send a program to the controller from the dashboard, and see which one is loaded right now
Drop a STEP file and its G-code into the browser to see the part and the toolpath together in 3D, with probe moves animated
Work-offset inputs, so you can check the program against the setup you actually have

STEPG-code3D toolpath

How it works

Measure, correct, observe.

One loop, running on the machine. The dashboard is where you watch it — not where it happens.

Data flow

01On the machine
Measure
A probe or measuring machine reports the finished dimension as soon as the cut ends. In the same moment the camera frame and the process sensors go to the inspection model.
02In the loop
Correct
The edge agent computes the wear correction, clamps it against the per-tool limit, and writes it to the controller before the next part is cut.
03Wherever you are
Observe
Every inspection, correction and alert replicates up to the dashboard. The edge holds a durable outbox, so a dropped link delays your data — it never loses it.

Only the edge agent sits on your network; the service itself runs in the cloud. The agent is built for NAT, outages and reboots.

Impact

What Quality 4.0 is worth

We do not publish other people's results as our own. These are published benchmarks for the category. What your line returns depends on your scrap rate, your cycle time, and how much of your inspection is still done by hand.

30–50%

Reduction in machine downtime

attributed to predictive maintenance in smart-factory deployments

McKinsey

10–20%

Reduction in cost of quality

across supply-chain technology programmes

Gartner

In-process

Not end-of-line

the shift from inspecting finished parts to correcting the process that makes them

IBM

Figures are industry benchmarks published by the sources named, not measured results from Undrd installations.

Platform

Fitted to the machines you run

Brands, controllers and vintages differ from one floor to the next, and so does what a machine will actually let you read. Rather than publish a compatibility list for you to check yourself against, we look at what you have, work out what it can expose, and build the integration around it.

How a machine gets connected

01
Survey what is on the floor
Which controllers are on the machines, what each one exposes, what is already wired, and how a part gets measured today. This is the step that decides everything after it, and it is where most of the surprises live.
02
Build the integration for it
Every controller sits behind the same fixed interface, so bringing in one we have not met before is a contained piece of work rather than a rewrite. Protocol, signals and timing get matched to the machine in front of us.
03
Close the loop on your line
Inspection, correction and reporting run against your tolerances and your cycle time. What the system does when a part drifts is a decision we make with you, not a default you inherit.

So far that has meant GSK, Fanuc and Renishaw hardware, over RS-232 DNC, Modbus TCP, FOCAS and in-process probing — on lathes, turning centres and machining centres. Yours may be none of those. That is a new adapter behind the same interface, not a new product.

Deployment

Edge agent on a Raspberry Pi or Revolution Pi, mounted beside the machine
Undrd Control is delivered as a managed cloud service — no server for you to provision, patch or keep alive
What leaves the plant is scoped with you before anything is switched on, rather than fixed by a default — your production data stays yours
Multi-instance from the start: one agent per machine, one dashboard for the fleet

Custom work

Custom development for what your factory actually does

No two shop floors are the same, and most of the interesting work sits in the gap between a platform and a plant. The system is built to be extended, not worked around with configuration.

New machine adapters

Controllers and measuring machines beyond the supported list, written against the same interface as the rest.

Models trained on your parts

Inspection models built on your defects — including one-class setups for lines where you only have images of good parts.

Line-specific logic

Gauging routines, sorting gates, traceability, and integration with the MES or ERP you already run.

Retrofit engineering

Wiring, I/O, edge hardware and the commissioning it takes to make it survive a real floor.

Undrd Control is delivered as a managed service, so there is no server for you to keep alive. What gets collected, where it is kept and who can see it are settled with you before anything is switched on — keeping your production data private is part of the engagement, not an afterthought.

Get in touch

Tell us about the line

The useful first conversation is a short one: what you make, what you scrap, and which controllers are on the floor. That is usually enough to say whether this fits.

Send us the details [email protected]