Reflections from our work in the Chips JU EdgeAI project — Digital Industry value chain
In most conversations about edge AI, the spotlight falls on the hardware: new chips, lower-power inference, tighter latency budgets. Far less often discussed is the quieter — and ultimately decisive — question of whether the people on the production floor will trust and adopt what we build.
Over the past months at Xtremion Engineering (XTE), this has been the centre of gravity of our contribution to the Chips JU EdgeAI project — a 48-partner, pan-European initiative co-funded by the Chips Joint Undertaking (Grant Agreement No 101097300), advancing intelligent processing at the edge across five industrial value chains. XTE contributes to the Digital Industry value chain, on two demonstrators that share a common industrial reality: automated defect detection on lines where human expertise and AI inference have to meet inside the same workflow.
Two demonstrators, one recurring question
VC1.5, led by SCM, develops edge AI quality control for the woodwork and furniture production process. VC1.6, led by HTS, addresses AI-based defect detection in System-on-Module (SoM) assembly, leveraging HD imaging from Automated Optical Inspection stations.
Both demonstrators bring together strong technical contributors — DEEPS, STMicroelectronics, University of Bologna, University of Cagliari, University of Milan — pushing the state of the art in edge-deployable vision models, explainability techniques, and distributed learning architectures.
Our role at XTE is deliberately different, and complementary. Edge AI systems in industrial environments are only as useful as the confidence and fluency with which operators, quality engineers, and line supervisors integrate them into their daily decisions. That confidence is not a natural byproduct of accuracy metrics. It is built carefully, through listening.
What the human-centred work actually looks like
Across the project we have designed and run structured user surveys, targeted questionnaires, and facilitated focus groups together with the technology partners who own the production environments where the demonstrators are being deployed. These sessions deliberately bring operators, quality engineers, and the development teams behind the AI solutions into the same room, in a format that lets unspoken assumptions surface early and be reconciled before they become design debt.
A few themes have emerged consistently across both demonstrators:
- Explainability is not a feature, it is a prerequisite. Operators want to see why a sample has been flagged as anomalous — not a probability score, but a visual cue grounded in the inspected object itself. This has direct consequences for how attention maps, saliency overlays, and confidence signals should be presented at the edge.
- False positives cost more than false negatives in the short term. A system that cries wolf erodes its own authority within days. The tolerance profile of the downstream process determines how aggressive the model can afford to be — and that is not a technical choice, it is an organisational one.
- Integration beats accuracy. A 99% accurate system that interrupts workflow is worse than a 95% system that fits it. Adoption is repeatedly described by users as a function of frictionless integration with existing inspection stations, reporting systems, and shift routines.
- Autonomy matters. Operators want AI to support their judgement, not replace it. The language of the interface, the granularity of override controls, and the traceability of every decision shape whether the system is experienced as a tool or as a supervisor.
Why this matters for European industrial edge AI
These findings feed directly into the requirements and refinement cycles of both demonstrators. They are also, we believe, part of what makes the European approach to edge AI distinctive. The Chips JU vision explicitly ties technological sovereignty to a human-centric, trustworthy, sustainable model of digital industry. Delivering on that vision requires that the people closest to the production process are not a downstream afterthought but an upstream input.
For XTE this is familiar ground. Our work in EdgeAI sits alongside our contributions to other Chips JU projects — H2TRAIN and EcoMobility — where we consistently hold the same position: at the interface between advanced sensing and semiconductor technologies, on one side, and the individuals and organisations that ultimately adopt those technologies, on the other. The hardest, and most interesting, problems in industrial AI are rarely only technical.
Grateful to the consortium — SCM, HTS, DEEPS, STMicroelectronics, UNIBO, UNICA, UMIL — for the collaborative spirit that makes this kind of work possible, and to the Chips Joint Undertaking for the strategic framework that enables it.
Blog signed by: XTE team