It is the question the electronics assembly industry has largely avoided asking out loud. Delvitech, the Swiss company that built its AOI and SPI platforms AI-native from day one, argues that the answer determines everything that follows: false-call rates, coverage of atypical components, and the true cost of owning inspection equipment.
By: Delvitech
A retrofit is not an architecture
Almost every supplier of automated optical inspection now describes its products as AI-powered, and in most cases the description is accurate. A neural network is present. It does useful work. What the phrase rarely reveals is where that network sits.
In the majority of systems on the market, the network was added to a decision architecture designed years earlier around rule-based thresholds. Deterministic algorithms measure a solder joint against tolerance windows and raise a flag; the network is then asked whether the flag deserves to be trusted. This is a genuine improvement and it produces a measurable drop in false calls. But the primary judgement is still made by rules, and the network never sees the evidence that produced them. It is a second opinion on a verdict already reached.
Delvitech began from the opposite end. Founded in Rancate, Switzerland, in 2018, the company specified its inspection platforms around proprietary neural networks from the first drawing rather than adding them to an existing product line. Horus, its all-in-one platform, performs AI 3D solder paste inspection together with pre-reflow and post-reflow AOI in true 3D. Aton addresses demanding post-reflow work, with multilevel inspection up to 140 mm. In neither case does the network act as a filter above a rule engine. It is the decision layer.
Where the intelligence sits determines what it is permitted to see
The architectural question is not abstract, because it decides the quality of the data the network receives.
A retrofitted network inherits whatever the original optics were designed to produce. If an optical head was specified for two-dimensional imaging with limited height sampling, no quantity of model training will recover information the sensor never captured. The intelligence is constrained by hardware that predates it.
Delvitech specified the sensor to feed the network. Its patented optical head uses phase shift profilometry, projecting sequential structured-light patterns through four synchronised projectors at 90-degree phase offsets. A single projector set measures features from a few micrometres up to 30 mm in height, which allows solder paste deposits and tall connectors to be measured on the same hardware. A working distance of roughly 300 mm keeps low-profile features adjacent to tall components measurable rather than shadowed. High-resolution 2D imagery and true 3D measurement are fused rather than sequenced, with the platform processing over 40 Gbit of inspection data per second, and a 50 by 50 mm field of view inspected in 0.6 seconds at 2 micrometer height resolution.
The last two per cent is where the cost concentrates
Contemporary inspection technology handles most standard production scenarios automatically, while the residual share comes from high variability, non-standard workflows, unusual components, and assemblies outside the training set. This is where costs accumulate through manual verification, rework, engineering hours and vendor callouts, precisely where retrofitted intelligence struggles most.
Delvitech’s answer is the Training Manager, which relocates the training lifecycle to the people who understand the process. Operators and process engineers label features directly, correct misclassifications and reintroduce them, and initiate training on the platform. Some training modes complete in under a minute. A single accurately labelled image can measurably improve model performance, and the tool is integrated into offline programming stations so that models are developed in parallel with live production and deployed in-line once validated, without pausing the line.
Operators can annotate and train it on-site to recognize and assess non-standard elements it initially cannot detect.
The problem is that the industry does not advertise
Any network that learns continuously faces a failure mode documented in the machine learning literature since 2013: catastrophic forgetting, in which a model trained on new data loses information it had already acquired.
Delvitech’s Knowledge Retention Technology is a continual learning method built to mitigate it. New image datasets can be absorbed without compromising previously acquired knowledge, allowing multiple models to evolve progressively rather than being rebuilt.
That knowledge is then portable. Validated models are distributed through a central library across every platform on a line, throughout a facility, and where required across geographically distributed plants, standardising inspection criteria and removing duplicated training effort between sites.
What the architecture pays back
The industry evidence for AI-enabled inspection is already substantial.
Delvitech’s claim is a behaviour, not a benchmark: because the neural networks interpret components and process variation rather than comparing them against static thresholds, false calls fall by a factor of ten. And keep falling, because every operator validation feeds the model. How far they fall on a given line is something Delvitech prefers to demonstrate on the customer’s own boards rather than assert in print.
The consolidation argument may matter more to a finance director. Running solder paste inspection, pre-reflow AOI and post-reflow AOI on one platform, one software environment and one programming environment means a single spares inventory, one operator training programme, and one set of recipes to maintain across the line. Setup-to-production time falls. Total cost of ownership falls with it. And because fewer false calls mean less unnecessary rework and less scrapped material, the sustainability case rests on the same mechanism as the financial one rather than on a separate claim.
Future-proof is a property of architecture, not a line in a brochure Self-programming and predictive functions, on the roadmap for 2026 and early 2027, are additions the architecture was built to accept. Retrofitting comparable capability onto a threshold-based system is not an upgrade path; it is a redesign.
Why the question matters here, and now
India’s electronics manufacturing base is scaling at a pace that makes this an operational question rather than a philosophical one. High-mix production, rapid product introduction and diverse component sets are exactly the conditions in which a fixed rule set ages fastest and an adaptable one compounds in value.
Made in Bangalore, to Swiss standards
Delvitech, recognized with the Swiss Economic Forum’s SEF.Growth High-Potential Label in 2025, is opening a manufacturing facility in Bangalore, alongside its Swiss headquarters and its US operation. At its core is a machine shop equipped with latest-generation high-precision CNC machining and inspection, producing critical components to the same standards as in Switzerland. Machines will be manufactured and assembled locally, with end-to-end testing, shortening lead times and delivery costs for Indian customers.
The facility will also serve as a platform for customer trials, technical evaluations, training and application support: the place where an Indian manufacturer can put its own boards on a Delvitech platform and watch. It is a long-term commitment to India, in the spirit of Make in India.
The decision facing a manufacturer investing in inspection today is not simply which system detects most accurately on the boards currently in production. It is which one will still be learning in five years, and who will own the knowledge it has acquired by then.
Delvitech will be at productronica India in Bangalore, 16–18 September, Hall 4 – Booth B21. The simplest way to settle the question this article opened with is to bring your most difficult board.
