Embedded AI & industrial GenAI
A model that runs next to a machine does not produce a dashboard, it produces a decision. What makes it usable is not its accuracy on a test set: it is what it is allowed to decide on its own, what it hands back to an operator, and what it does when the scene falls outside anything it was trained on.
The hard part is not the model. It is the boundary you give it.
Training a model has become the short part of the work. What decides whether it goes into service is what it is allowed to settle on its own, what it hands back to an operator, and what it does when the scene falls outside what it knows.
We write that boundary with your operations teams before touching the hardware, then make the model fit the equipment already installed on site. That is what separates a convincing demonstration from a model a shop floor agrees to let decide.
What we do
Four workstreams, from the boundary written for the model to the fleet that carries it.
Write what the model decides alone
What it settles, what it never settles, the band of doubt it is allowed to hand back, and who arbitrates it. The document is signed by operations, because stopping a line commits people's safety and that responsibility stays human.
a model whose perimeter is unwritten ends up deciding everything
- What it decides, and what it never decides
- A band of doubt handed to an operator, not settled
- Every automatic decision replayable for twelve months
Make it fit the hardware already on site
Quantisation, pruning and compilation for the gateway that is already installed, with a latency budget taken from the process rather than from a benchmark, and the accuracy given up to compression measured and accepted in writing.
a size gain never priced in accuracy is discovered in production
- Latency budget taken from the process interval
- Accuracy before and after compression, both written
- Previous version kept locally for rollback
Copilots grounded in the shop floor
Procedures, maintenance history and equipment documentation turned into an assistant whose search is restricted to validated sources, whose answers cite where they come from, and which stays short of anything touching the physical process.
an assistant that cannot cite its source is not usable in operations
- Search restricted to validated sources
- Answers citing the procedure they come from
- Human in the loop on anything touching the process
Operate a fleet of models
Rollout in waves starting with one pilot site, over-the-air updates, weekly comparison of score distributions against the acceptance set, and retraining fed by what the field actually sends back.
a model suddenly flagging twice as much means the line changed
- Rollout in waves, one pilot site first
- Drift watched against the acceptance distribution
- Field images feeding the next training round
Where the decision runs
A decision taken within the interval the process leaves, without a variable round trip.
A round trip that is fine for a report, and not for a part already gone past.
What leaves the site is a result, not the raw scene it was read from.
The raw data has to be sent, and therefore held, wherever it is processed.
Keeps deciding while the link is down, and reconciles once it is back.
A dropped link stops the decision, and the process does not wait.
What you get
One project runs through the four deliverables below: a detection model deployed on a fleet of gateways. Each line states what is actually handed over, in the order it is handed over.
A written perimeter of decision
What the model settles alone, what it never settles, the band of doubt, who arbitrates it and how long a decision stays replayable. Signed by operations, not by the data science team.
A model packaged for the gateway on site
Quantisation, pruning and compilation for the hardware already installed, with the latency budget taken from the process and the nine thousandths of accuracy given up written down rather than discovered.
Guardrails that can answer nothing
Out of its operating domain the system does not answer badly, it does not answer at all. The band of doubt goes to an operator with the scene, and drift is compared weekly against the acceptance set.
An evaluation read by condition
Latency and footprint hold, overall accuracy reaches 0.938, and it splits into 0.961 by day against 0.857 on the night shift. That gap is what decides how the model goes into service.
How we deliver
Discover
depending on scope, sector and the level of compliance required
- Audit of use cases and pain points
- Value / feasibility matrix
- Executable specification (ASDD)
MVP
depending on system complexity and integrations
- An agent in a real environment
- Generated tests, measured coverage
- Go / no-go before industrialisation
Scale
depending on the number of agents and connected systems
- Multi-agent orchestration on a shared foundation
- CI/CD and MLOps integration
- Team upskilling
Run
service commitment defined with you
- LLMOps observability
- FinOps optimisation of AI costs
- Continuous compliance audit
Where industrial AI actually stands
Models that decide in production
Inference on the cameras, no biometric data leaving them
142 4K cameras · p99 latency under 200 ms
Fifteen million visitors a year whose flows had to be read as they happened, on a site where sending video streams away for central analysis would have created exactly the personal data nobody wanted to hold.
Inference run on the cameras themselves rather than on the footage: what leaves is a count and a signal, not an image, so the flow is managed in real time and no biometric data is produced at any point in the chain.
A reading handed back to a practitioner, not a score
96% model accuracy · error rate 35% → 1%
Dental charting done by hand on panoramic X-rays, slow and uneven, where the model's output is not a dashboard figure but a clinical reading a practitioner will act on immediately.
Deep learning built into the product and co-designed with the practitioners who use it, with accuracy treated as a delivery criterion, and a hosting and privacy frame set for health data from the start.
Insights & Perspectives
Context engineering: how to give AI exactly what it needs
Skeleton trimming, relevance-based file selection and progressive contextualisation: three techniques for models that answer on the right material rather than on everything.
Why the four key metrics matter in a software-driven automotive industry
Adapting the four DORA metrics to software-defined vehicles, through simulation, hardware in the loop, walking skeletons and vertical slicing.
The age of hyperautomation
Combining robotic process automation, AI, process management and analytics to automate at scale: components, platforms, benefits and a deployment method.
Put a decision next to the asset, with its boundary written
A perimeter signed by operations, a model that fits the hardware already on site, guardrails that can answer nothing, and an evaluation read by condition.
Frequently asked questions
Because of what leaves the site as much as because of speed. Inference on the camera sends out a count and a signal instead of the footage, so the personal data is never produced in the first place. The interval a process leaves between two parts is often shorter than a round trip, and a link that drops must not stop the decision.
Some accuracy, and the amount has to be measured rather than assumed. On a typical detection model, quantisation to eight-bit integers and channel pruning cost a few thousandths, which is acceptable when written down and accepted. A size gain never priced in accuracy is a gain you discover in production.
Nothing, and that has to be built deliberately. A model never says it does not know: it always returns a score. The code around it refuses to ask when the operating conditions are out of range, and hands the band of doubt to an operator together with the scene, because a score without the image is an instruction rather than an aid.
It can quote the procedure and say where it comes from. It does not interpret it, and it stops short of anything touching the physical process. Search is restricted to validated sources, every answer cites the document it came from, and the interaction stays auditable for operations and safety teams.
The timing changed in July 2026. Regulation (EU) 2026/1744, published on 24 July 2026, moved high-risk obligations for standalone Annex III systems to 2 December 2027, and to 2 August 2028 for AI embedded in a regulated product under Annex I. The extra time is not idle: risk management, documentation and human oversight all have to be built into the product.
By comparing the distribution of its scores each week against the one measured at acceptance. A model that starts flagging twice as much has not learned anything: the line changed, the lighting changed, or the parts did. The response is to widen the band of doubt while the cause is found, not to move the threshold.
Framing takes 2 to 6 weeks depending on scope, sector and the level of compliance required, and produces the perimeter of decision, the latency budget taken from the process and the target hardware. A first model evaluated on a pilot site, under real conditions rather than on a test set, follows in 4 to 10 weeks.