Stellantis Goiana: DataOps and UNS Behind 30+ AI Applications

On 26 October 2026, IIoT World published an article describing how Stellantis' Goiana plant in Pernambuco, Brazil, built a scalable data architecture on top of its existing Siemens infrastructure. The architecture now supports more than 30 AI applications. According to the account, integration times dropped from months to weeks.

A note on the source: the article is sponsored by HighByte (labeled "Editorially Independent") and is based on a session at the Industrial AI Summit 2026, with speakers from HighByte and Mekatronik. The results should therefore be read as vendor-reported. No costs are given, nor details on the individual AI applications.

Why the dashboard approach is no longer enough

For years, PLC tags fed dashboards and OEE analytics. With machine learning and digital twins, requirements change: more sources are needed, such as SQL databases, CSV files, images and video. Building a dedicated integration for every application across hundreds of machines does not scale.

The contrast, as presented, is stark. The "dashboard" architecture relies on application-specific calculations and manual integrations for each line and plant. The "AI-ready" one contextualizes raw data once for all consumers: a model is built once and replicated across lines and plants.

Industrial DataOps and the Unified Namespace

According to the speakers, DataOps covers about 80% of the data preparation work for AI: connecting to shop-floor sources and adding context. The contextualized data is then published through a Unified Namespace (UNS).

Aron Semle (HighByte) noted that the UNS is implemented in different ways: in an MQTT broker, a data lake, a historian or an OPC UA server. The common element is not the technology but the consistency and contextualization of the data.

Goiana's two-layer model

The architecture described is organized as follows:

• Lower layer: handles the protocols and formats of each machine.
• Upper layer: maps data into a common model, so machines with the same function on different lines produce the same data structures.

The upper layer took more effort. Because this is discrete manufacturing, it was not enough to attach meaning to tags as in process industries: the team had to model the parts moving through the line, that is, operations, sequences and materials. Once working on one line, the model reaches the next much faster.

PLC data accounts for about 80% of the volume, but the decisive context comes from the bill of materials, product details and cycle data.

Other implementation timelines cited

The account also reports other examples, with the caveat that timelines depend on each site's processes, systems and documentation:

• a plastics plant connected its machines and started statistical process control within a week;
• a process site with thousands of tags on another system brought its first applications online within a few weeks, while the full migration took a few months.

The vendor also cites more than 400 sites in 25 countries for its software and an increase in mechanical availability from 70% to 95% in six months at Gousto. The latter figure is unverified: the article does not explain how it was measured.

Key takeaways

The main message is that the bottleneck is the data model, not the AI model. For those planning AI initiatives on the factory floor, a few practical pointers:

• define a common data model for equivalent functions before multiplying use cases;
• in discrete manufacturing, include operations, sequences and materials in the model, not just tags and units;
• treat the UNS as an outcome (consistent, contextualized data) and choose the technology to fit your own context;
• validate on one line, then replicate, measuring real integration times.

The costs and benefits of the individual applications remain to be verified with independent data. Even so, Goiana is a useful reference for what a UNS needs to cover in a discrete environment.