Rio Tinto builds AI assistant documenting Metpro system

Rio Tinto built an AI domain assistant to capture and expose the embedded knowledge, dependencies and decision logic inside a 30-year-old manufacturing execution system, Metpro, which manages the aluminium product lifecycle for its Australia and New Zealand aluminium operations, Ke Shi, data science senior advisor, said at the AWS Summit Sydney. Ke Shi reported that Metpro's technical documentation had become fragmented across "thousands" of documents and that undocumented coupling meant small changes could have unexpected downstream impacts.

Ke Shi said Rio Tinto created a "domain-aligned training dataset" by combining the Metpro code base with "business and operational context" and ingested that into Amazon Bedrock Knowledge Bases and Amazon Bedrock AgentCore. The team used Amazon SageMaker AI Jumpstart to evaluate foundation models and adopted Llama 3.1 8B as the initial inference model, according to the AWS Summit remarks.

Companies integrating foundation models with operational technology often aim to centralise fragmented documentation and make system behaviour discoverable without replacing legacy platforms. Observed patterns in comparable deployments include the need to manage data lineage, guardrails for model outputs, and integration with change-management and safety workflows.

metrics on onboarding time, incident/change-related outages, versioning and audit trails for the assistant, and how the organisation vets and updates the training dataset and model choice over time.