AI on a Zimbabwean mine: from cameras to decisions

Sector figures

Three use cases with risk notes

1. 01

   Vision at intake, gold room and perimeter

   Cameras with edge inference flag events (person in a restricted zone, object in a screening lane, vehicle at a gate) and push them to security and control-room staff who decide. The model runs on site; the site keeps working when the link or grid does not.

   Risk notes. The moment the system identifies a person, it is processing biometric data: written consent required (Act s.12(1)), processing notified to the Authority (SI 155 s.10(2)(d)), and no decision that affects a worker taken solely by the system (s.25). Event detection without identification carries far less exposure and delivers most of the value. Related company Eigenstate Systems works in this physical-intelligence domain.

2. 02

   Predictive maintenance from the plant historian

   Models read vibration, temperature, current and throughput series from the historian through a one-way tap and rank assets by failure likelihood for the maintenance planner. No write path to control systems exists.

   Risk notes. Operational data, little personal data; model risk is a production and safety matter, so validate against outcomes before the planner relies on it; historian tap must be read-only and network-segregated from OT; document the model in the register even though no RBZ standard applies, because insurers and auditors will ask.

3. 03

   Grade control and blend decision support

   Models combine assay, geology and plant data to recommend blend and dispatch decisions; a metallurgist or mine planner approves. Recommendations and outcomes are logged so the model can be measured against recovered grade.

   Risk notes. Commercially sensitive data; keep in-country; outcome analysis and periodic re-validation; clear ownership between geology and metallurgy for the model; the recommendation is never executed automatically.

Power, links and where inference runs