From Cloud to Factory Floor: The Evolution of MLOps Architectures for Manufacturing

TL; DR;

"MLOps started in the cloud. Now it must survive on the factory floor."

Google’s canonical MLOps architecture set the foundation for ML systems engineering — defining CI/CD pipelines, model registries, automated retraining and performance monitoring. It works well for cloud-native apps, SaaS platforms, and enterprise IT.

But on the industrial shop floor? It struggles to survive. Let's dive into this topic in this article.

The gap between IT and OT

Industrial manufacturing lives in a different universe:

- 🔌 **Intermittent or no cloud connectivity**
- ⏱️ **Sub-second latency for critical systems**
- 🔐 **Sensitive data locked behind OT firewalls**
- 🧱 **Legacy equipment using OPC UA, SCADA, and PLCs**

These constraints make the cloud-first architecture unfit for real-time control loops, predictive quality models, and embedded diagnostics.

A new architecture is emerging: hybrid edge-cloud

A recent scientific publication by Raffin et al. (2022, FAU Germany) proposes a reference architecture tailored to manufacturing. Unlike the centralized Google model, this architecture is **distributed, modular and event-driven**.

It’s not about abandoning the cloud — it’s about **empowering the edge**. Here’s how:

🧠 Inference at the edge

Models are served locally via lightweight containers, with inference times in the order of milliseconds. This ensures safety, uptime, and operational continuity even when offline.

🧩 Microservices and domain-driven design

Each pipeline component — from data acquisition to decision services — is a microservice deployed on the edge. Interfaces are designed for modularity, encapsulation, and adaptability to factory heterogeneity.

 🔄 CI/CD/CT adapted to OT

A full DevOps loop is preserved, but adapted:

- Models are versioned and tested centrally
- Containers are deployed locally using GitOps or IaC
- Continuous Training (CT) is triggered by drift or scheduling
- Monitoring is embedded both locally and in the cloud

 📊 Governance and feature stores

Data is cached and served via local feature stores. Model drift, data quality and inference logs are managed on-site, with batch synchronization to cloud systems for retraining and collaboration.

Why it matters

This shift isn’t just technical — it’s operational and strategic. The success of industrial AI depends on:

- Running models where the data lives
- Empowering process engineers, not just data scientists
- Handling millions of edge nodes without losing control

The hybrid architecture enables AI to operate autonomously, securely, and explainably — without sacrificing agility or compliance.

How aignosi is leading this transition

At aignosi, we’ve implemented these principles in our platform — combining:

- 🔹 Industrial-grade data collectors (OPC UA, MQTT, Modbus)
- 🔹 Scalable edge-based orchestration
- 🔹 Embedded MLOps governance and observability
- 🔹 No-code/low-code interfaces for domain experts

With our solutions, manufacturers are running dozens of AI models per site, even in air-gapped or partially connected environments.

We believe the future of AI in manufacturing is not cloud-only — it’s cloud-aware, edge-native, and operator-centric.

---

📚 Reference: Raffin, T. et al. (2022). A reference architecture for the operationalization of machine learning models in manufacturing. Procedia CIRP, 115, 130–135.*