Carolina

Co-founder @IIoT_World | Industrial AI, IoT & Cybersecurity. Helping leaders share what works & what's next. Join us for the 2026 World Days Series!

[Carolina @CRudinschi]

Scaling AI across equipment, plants, and fleets requires continuous model tuning using real operational data. That's why data-sharing infrastructure and commercial models between OEMs and operators must be in place before scaling begins.

[Carolina @CRudinschi]

An AI vibration model that only sees a spike in isolation raises false alarms. A model that also knows the product being run, the line speed, and the maintenance history can distinguish genuine faults from normal process variation.

[Carolina @CRudinschi]

Continuous feedback from operators into the model is what reduces false positives over time. A pilot lacking this feedback loop stays at pilot quality and never earns the trust production teams need to act on its alerts in mission-critical situations.

[Carolina @CRudinschi]

Source: https://t.co/ffGz4RdWtn

[Carolina @CRudinschi]

In a conventional point-to-point OT architecture, an external AI app can connect directly to a PLC via OPC UA and write a new setpoint. Mekatronik's DataOps layer intercepts that, validating range and operating state before any value reaches the control system. #highbyte_iiot

[Carolina @CRudinschi]

With an infrastructure-centric model, organizations can add AI use cases incrementally instead of treating every new project as a greenfield infrastructure effort, which reduces both the time and cost of each additional deployment.

[Carolina @CRudinschi]

Every new AI project in a point-to-point integration environment requires building a fresh set of connections from scratch. A predictive maintenance pipeline, then an energy optimization feed, then a quality monitoring link. Each one adds its own maintenance burden. Partner content with Cybus. #hivemq_iiot #cybus_ai

[Carolina @CRudinschi]

The alternative to point-to-point integration is an infrastructure-centric approach: a shared data platform connects once to all operational sources and makes that data available to any application through standardized interfaces.

[Carolina @CRudinschi]

How long before an edgeRX sensor gives you useful data? Vibration and temperature readings are immediate. Full AI anomaly detection kicks in after two weeks of operational baseline. Tim Gicewicz, TDK SensEI at IMTS. Partner content with TDK. #TDK_iioT

[Carolina @CRudinschi]

Boris Scharinger of Siemens makes a similar point about solution providers: those who have already built and deployed diagnosis-to-therapy logic will typically help plants achieve the desired outcome faster than teams building the stack from scratch.

[Carolina @CRudinschi]

The CEO of Infinite Uptime describes the architecture in three layers: physics-informed fault signatures handle initial classification, a human expert verifies prescriptions, and RAG+LLM sits on top as an explanation and search layer, each layer doing what it does best. Partner content with @AI4ProdOutcomes. #infinite_ai

[Carolina @CRudinschi]

Source: https://t.co/431fnRPVh7

[Carolina @CRudinschi]

A traditional cyber risk register models threats from adversaries. An industrial AI risk register must capture authorized systems making bad decisions with physical consequences, a failure mode cyber registers are structurally blind to.

[Carolina @CRudinschi]

The most likely AI failure in OT involves no adversary at all, just an authorized system operating as designed, making a bad decision with real physical consequences. Matt Morris of EverLine explains that industrial AI risk belongs closer to process safety than to cybersecurity. Partner content with @skkynetinc. #skkynet_ai #everline_ai

[Carolina @CRudinschi]

Morris recommends running industrial AI risk like a HAZOP, applying guidewords more, less, none, reverse, late, other than to every AI action and asking what happens. Industrial AI registers must ask what if it works exactly as built and is wrong.

[Carolina @CRudinschi]

Audience questions at the Industrial AI Summit 2026 session covered five topics: edge AI deployment, AI platform selection, LLMs in production planning, digital twin validation, and heterogeneous data contextualization, reflecting how broad the practical implementation gap still is.

[Carolina @CRudinschi]

At the Industrial AI Summit 2026, manufacturers asked whether AI could be validated on a digital twin before going live.

[Carolina @CRudinschi]

An audience member at the Industrial AI Summit 2026 asked whether domain-specific small language models trained for a particular process could run at the edge to reduce latency and security risks while sending only aggregated KPI data to the cloud for monitoring.