Domain
Digital & AI
Intelligence that has been asked to do a job, not merely to impress a slide.
We are fond of models that can be explained to a sceptical plant manager. The mathematicians are here. Industry has the mess. Introductions follow.
The useful question in industrial AI is almost never which architecture. It is what decision changes as a result, who is accountable when the model is wrong, and what the two kinds of error cost. A false positive on an inspection line costs a teardown. A false negative costs rather more, later, in front of a customer.
This makes calibrated uncertainty more valuable than accuracy. A model that reports low confidence and hands the part to a human is deployable. A model that is right ninety-six per cent of the time with total confidence is a liability, and everyone on the shop floor works this out faster than the project board does.
Data is the other reliable disappointment. Industrial datasets are small, imbalanced, drifting, and labelled by whoever had time. Projects that budget seriously for labelling and for a held-out set collected on a different week generally work. Projects that do not tend to produce a demonstration that only ever worked on the training set.
- Industrial AI that respects a safety case
- Digital twins that are twins, not cousins
- Data governance that would survive a Select Committee
- Uncertainty reported honestly, and acted on
- Models that a quality engineer can interrogate
What sits inside
The sub-disciplines people actually search for.
Process monitoring and machine vision
Defect detection, in-process sensing, and the discipline of collecting a test set on a different week from the training set.
Digital twins and simulation
Calibrated models with a stated validity range, rather than a rendered façade with a live data feed attached.
Scientific machine learning
Surrogate models, physics-informed methods, and inverse design for problems where a full simulation is too slow to be useful.
Assurance and governance
Test strategies for systems that learn, documentation a regulator will accept, and a clear account of what the model was never trained to see.
Edge and embedded deployment
Inference on hardware that lives in a cabinet, tolerates dust, and cannot be rebooted during a production run.
Where the UK stands
Without the boosterism, and without the gloom.
- The research base is genuinely strong and unusually well distributed — this is not a single-city domain.
- The Alan Turing Institute, the Hartree Centre and the Catapult network provide most of the routes from paper to plant.
- Compute access, not talent, is the common bottleneck for mid-sized industrial problems.
- Assurance expectations are rising sharply in regulated sectors, which favours teams who documented properly from the first sprint.
Work that gets commissioned
The shape of a package in this domain.
- Feasibility studies on a real, messy dataset with an agreed success metric
- Model development with calibrated uncertainty and a defined hand-off to a human
- Independent validation of a supplier's model on data it has not seen
- Digital-twin calibration against instrumented plant measurements
- Assurance documentation for a system entering a regulated environment
Where projects come unstuck
Reliably, and usually in week nine.
- A demonstration that only works on the training distribution, discovered in week nine
- Reporting accuracy on a balanced set when the real base rate is one in four hundred
- Building an interface for a dashboard nobody has agreed to look at
- Deploying without a drift check, then wondering why performance decayed after a supplier change