Data & AI.
AI Engineering in practice.

The operating problem
Lab assays describe a composite sampling window and arrive later. Joining them as point-in-time readings creates misleading recovery labels.

The practitioner approach
Historian-to-warehouse ingestion with assay alignment, plant-stop validity handling, sensor health checks and soft-assay models.

Why it matters
Recovery visibility depends on the data’s timing and physical meaning, as well as the pipeline.

Relevant starting point
Recovery Readiness or Model Validity Audit.

The operating problem
Grinding and flotation interact. Improving one stage can reduce performance elsewhere, and recovery must be considered alongside concentrate grade.

The practitioner approach
Stage-level balances and models connected under a plant-wide objective, with operational bounds and metallurgist-defined physical constraints.

Why it matters
A useful optimiser accounts for equipment state, permissible step changes and the propagation of upstream effects.

Relevant offering
Recovery visibility, progressing to operator-reviewed set-point advisory.

The operating problem
An accurate model can remain unused if its recommendations do not earn the control room’s trust.

The practitioner approach
Metallurgist sign-off, simulator sessions, recommendations framed as changes from current state, and captured acceptance and rejection reasons.

Why it matters
Adoption requires checking what changed in the plant, as well as what operators clicked. Advisory stays outside the control system.

Relevant offering
Model Validity Audit, validity monitoring and advisory design.

The operating problem
Operators know about unfamiliar ore or equipment changes that sensors alone may not reveal.

The planned extension
An AI reasoning layer would read plant context and prior rejection reasons to adjust the next recommendation before operator review.

Status
Designed, not completed at the reference engagement. A production feedback loop is not claimed.

Relevant offering
A future extension to operator advisory, after visibility and monitoring are established.

The problem
Customer enquiries arrive by email and need resolution across business systems.

What was built
A support workflow that resolves emails end to end and records its decisions for audit.

Current status
In production. The agent graduated from human review.

Reference scope
Engineering example. Measured business outcomes are not published.

The problem
Preparing quotes requires matching products, applying pricing and reviewing revisions.

What was built
An invoice-to-quote workflow producing matched, priced and versioned quotes with four account-manager gates.

Current status
Implemented. Broader workflow development remains in progress.

Reference scope
Engineering example. No measured savings are claimed.

The operating problem
Shipment paperwork arrives across multiple documents and days. Re-keying it is manual, and conflicting values require broker judgement.

What was built
An intake pipeline that classifies documents, extracts sourced fields and reconciles values. Each field retains its source document and page.

Current status
In production for customs entry filings. A licensed broker reviews and submits every draft; the AI has no filing path.

Reference scope
An anonymised engineering example. Extensions to other filing types are scoped; no measured business savings are claimed.