Systems that work on the line, not in the pilot cell.

Manufacturing AI fails physically far more often than it fails mathematically: a moved camera, a changed lamp, a new supplier's material, a sensor nobody recalibrated.

01 / What is hard here

The problem underneath the problem.

A pilot runs in a controlled cell with good lighting and a cooperative operator. Production has three shifts, vibration, changing daylight through a roof panel, and a maintenance team who will reposition your camera without telling anyone.

The defects you most need to catch are also the rarest. On a line with a 0.2% defect rate, a model that declares everything acceptable is 99.8% accurate — and an untended training process will find exactly that solution.

02 / What we build

Work we take on in manufacturing.

Visual inspection: Surface, assembly and dimensional checking with anomaly-based methods where defect examples are too scarce to train a classifier, deployed on hardware at the line.
Predictive maintenance: Condition monitoring from vibration, current, temperature and controller logs, with a warning horizon designed around your parts lead times and maintenance windows.
Yield and scrap analysis: Tracing quality variation back to process parameters, shift, batch, material lot and machine — usually the fastest route to a return in a plant with historian data.
Production scheduling: Sequencing under changeover, capacity and due-date constraints, producing a plan a supervisor can adjust rather than a black-box instruction.
Energy and utilisation: Consumption modelling per line and per product to locate the difference between nominal and actual efficiency.

03 / Constraints

What shapes the build.

These are design constraints agreed at the start, not compliance work bolted on before launch. Retrofitting them is expensive and usually incomplete.

The capture path is part of the system: Camera placement, lighting, mounting and trigger timing specified and documented. Most field failures we are called in to fix are physical.
Legacy equipment: Machines from different decades with proprietary or absent interfaces. Getting data out is frequently the largest single line item.
Network independence: Inference runs at the edge so that a connectivity outage does not stop the line.
Operator trust: A system with two false alarms per shift will be switched off. The false-alarm budget is agreed with the line before tuning, not after.

04 / What we advise against

Where we would say no.

We would advise against a vision project where the defect cannot be seen under achievable lighting, and against predictive maintenance where maintenance history is not recorded in any recoverable form — that is a data-collection phase first, and it should be budgeted as one.

We would rather lose the work than build something that creates a liability for you or a harm for the people on the other side of it. If that rules out what you had in mind, it is better established now than at delivery.

05 / Other sectors

Adjacent work.

Different domains, the same delivery discipline — a measured baseline, a real evaluation set, and monitoring that catches decay early.

Next step

Working in manufacturing?

Tell us which process is slow, expensive or error-prone. The diagnostic will tell you whether this is a problem worth solving with machine learning — including if it is not.

Request a diagnostic All industries