Co-development R&D and proof of concept
Applied research run with your engineers, on your process, against a target both sides agree on before the first recording is made.
Most industrial AI projects fail in the framing, not in the model
Most industrial AI projects do not fail in the model. They fail in the framing: nobody wrote down what success would look like, the data was collected once and never checked, and the evaluation flattered the method instead of testing it. By the time the results are presented, the plant has lost six months and the engineers have lost interest.
We run R&D the way a measurement campaign is run. The problem is stated in the plant's own terms, the success metric is a number with a threshold, the data is designed to cover the process window rather than one good week in summer, and the model is judged against baselines that are hard to beat. If the physics is not there, you hear it from us early and cheaply, which is a better outcome than a prototype nobody dares deploy.
The work happens with your process engineers, quality engineers and maintenance staff in the room, not afterwards. That is not a courtesy; it is where the label definitions, the awkward edge cases and the real operating conditions come from.
What is included
Joint problem definition and success metrics; measurement plan; dataset engineering (label specification, QA, versioning); representation and model design; evaluation under realistic operating conditions with strong baselines; error analysis; production plan. Suitable for internal R&D budgets, publicly funded projects and university collaborations.
For German clients the follow-on R&D is usually structured as a ZIM co-operation project, which recovers part of the client's eligible cost; we prepare the technical sections.
A measurable target, written down
Detection rate at a fixed false-alarm rate, or scrap reduction, or cycle-time budget. Agreed in week one, reported against every milestone.
A dataset you own
Measurement plan, label specification, QA and versioning, documented so your team can retrain on it in three years without us.
A model evaluated honestly
Strong classical baselines, held-out data from different days and variants, ablations, and an error analysis that names the cases it still gets wrong.
A production plan or a clean no-go
Hardware class, integration points, expected unit cost and the risks that remain — or a written recommendation not to proceed, with the evidence.
Four stages, each with an output you can check
Problem definition workshop
Half a day with your engineers: the failure mode, the current method, the cost of getting it wrong, and the number that would make this worth doing.
Measurement and data design
Sensor choice and placement, sampling rates, the process window to cover, and the label specification written with your quality engineers.
Modelling and evaluation
Representation design, baselines, learned models, evaluation under realistic conditions, error analysis with your team.
Decision gate
Written report, live demonstration on held-out data, and a costed proposal for the pilot — or an honest no-go.
Sectors that buy this most often
What clients ask before they start
How is this different from hiring a data science consultancy?
A general consultancy will bring modelling skill and no sensing. The hard part in this domain is upstream of the model: which physical quantity carries the signature, where the sensor goes, what the label actually means, and how the signal changes when the line speeds up. That is what we do, and the modelling follows.
Can we start with a proof of concept before committing to R&D?
That is the normal path. A two-to-eight-week proof of concept on one asset or one station costs a fraction of a full project and produces the evidence you need for the internal business case.
Who owns the intellectual property?
The dataset, the trained weights and the documentation are yours. Background methods stay ours. Anything we want to publish is agreed with you first, and nothing identifiable is published without written permission.
Ask about Co-development R&D and PoC
Tell us the asset or process and what goes wrong. We answer every enquiry within 48 hours on working days, and the first call is free.
- NDAs signed before the first call if you prefer
- Anonymised or synthetic samples are fine to start
- Your data is never used to train models for anyone else
Services clients combine with this one
Ready to find out if it works on your machines?
Start with a free 30 to 45 minute discovery call. Bring the problem, the recordings you already have and your questions. We will tell you honestly whether sensing-based AI is the right tool and what the next step would cost in days, not months.