CodeNinja Atoms has published an open industrial AI reference architecture for monitoring high-consequence tailings storage facilities, offering a practical blueprint for deploying physical AI on operator-owned hardware. The documentation, authored by Muhammad Umar and available under CC BY 4.0, details a system where AI agents draft insights but named engineers retain final approval authority. This approach rejects the trend of cloud-dependent model services in favor of a governed ontology that sits beside existing systems of record without replacing them.

Architecture Prioritizes Data Sovereignty

The design enforces strict data sovereignty by keeping all instrument, survey, and inspection data on the operator's hardware within the United States. No third-party model services are called, and the system avoids fine-tuning on trigger history or using frontier clusters. Instead, it relies on a three-tier inference stack running on local infrastructure: edge units use Granite Tiny Time Mixers for screening, while the site server room employs Chronos-2 for trend baselines and Nomic Embed Text v2 for retrieval. The drafting tier utilizes gpt-oss-120b served by vLLM on a node equipped with eight 80 GB GPUs, ensuring that sensitive industrial data never leaves the facility.

Object Model Connects Disparate Systems

At the core of the architecture is a governed object model built on Hyper Ontology v0.9, which defines fifteen objects and twelve typed links to connect eight disparate data sources. These sources include instrumentation historians, satellite radar reports, and regulator correspondence, all accessed via read-only adapters that enforce a no-modification rule. The model allows engineers to query a single embankment section and retrieve a comprehensive narrative by traversing links between readings, raise events, deposition records, and inspection observations. This structure solves the problem of fragmented data, where critical patterns are often hidden among single-instrument alarms in traditional historians.

Human Verification Remains Central

The workflow explicitly separates AI drafting from human confirmation to maintain safety and accountability. Three agents draft pattern, response, and reporting insights, but every draft must be confirmed by a tailings engineer or approved by the engineer of record before reaching regulators. This human-in-the-loop mechanism ensures that AI outputs are validated against engineering judgment, with all confirmed patterns stored as memory for future reference. The system does not train models on the fly; instead, it relies on retrieval-augmented generation and static models to provide context, keeping the operational logic transparent and auditable.

Key Takeaways

  • The reference architecture is free to reuse under CC BY 4.0, including the paper, object model JSON, and model register.
  • Inference runs on three tiers of local hardware, using Apache-2.0 licensed models like gpt-oss-120b and Chronos-2.
  • A governed ontology joins eight data sources without modifying the underlying systems of record or requiring a data lake migration.
  • Human engineers retain final approval authority on all AI-generated drafts before they are submitted to regulators or actioned.
  • The system costs approximately $456,000 to own over three years, significantly cheaper than renting equivalent AWS capacity while maintaining data sovereignty.

The Bottom Line

This architecture proves that industrial AI doesn't need to sacrifice data sovereignty for intelligence, offering a cost-effective, auditable alternative to cloud lock-in that puts human engineers back in the driver's seat.