Kopala Compass

AI mineral intelligence / Zambia

Find copper.
Dig safely.
Stay legal.

One AI engine turning Zambia's geological data into ranked exploration targets.

Zambia Challenge 1: From Geological Information to Exploration TargetsMineTech Hub Pre-Incubation Programme · Mukuba UniPod, Copperbelt University
Built by Michael Chileshe, cloud engineer and AI developer, and Henry Chemba, nuclear physicist

12°45′S / 28°03′E

One engine / Explorers first, miners in phase 2

Intelligence that travels
from boardroom to borehole.

01Where to look

Every square kilometre ranked by copper potential, from new airborne survey data, legacy geological records and satellite imagery.

02Is it legal

Company licences, other miners' licences, forest reserves and riverbank zones checked before anyone digs.

03Safe tonight

Changing risk from rain, slopes, pits and waste dumps, sent to licensed miners by SMS before the night shift. Advice only, never a clearance.

Copperbelt / Pilot workspace

Exploration Dashboard

Forecast

Exploration intelligence / Zambia

Copperbelt prospectivity

12°45′S28°03′E1 km grid

Model output / Copperbelt pilot

Ranked Targets

12 new targets

Rank	Type	Score	Area km²	Sand cover m	Can it be worked?	Confidence	Verification	Location / Risk tonight

Phase 2 / Field channel / USSD + SMS

The same targets on any phone.

Licensed artisanal miners dial a short code to check their licence area: copper potential, whether it is legal, and whether it is safe tonight. No smartphone or data needed.

Target report / 01

Priority target

Model score99/100

Model interpretation

Why the model
chose this.

Next action

Verify before you mobilise.

Expert verification Field notes Saved in this browser. The production version shares reviews with the whole team.

Evidence / Region hold-out on real deposits

Does the model find ore
it has not seen?

Success-rate curve / Higher and further left is better

How much ground you must search.

Same folds, four model types / Two tests

Model comparison.

Model	Hold-out: top 10%	Strict blind: top 10%	Map for 80% (worst)

Is the test rigged?

Null tests.

Test	Found in top 10%	Should be

Each region hidden in turn / Lower is better

Where each hidden deposit ranked.

Top % of the map · dashed line 10%

Permutation importance

What the model relies on.

Honest limits

Where it is weaker.

Situation	Found in top 10%	Median rank

Data integration / One 1 km grid

What is real, what is a stand-in.

Layer	Real source	Type	In this prototype

Model

Labels: USGS MRDS

Phase 2 / Field channel / USSD + SMS

Critical intelligence.
No data connection.

Exploration guidance for licensed miners, and safety checks for anyone at a known digging site, even without a licence. Any basic phone, in English, Ichibemba or Chinyanja. Copper scores are never shown for ground a caller may not dig.

182 Character screens17:00 Nightly alert5 GPS checks a day3 Languages

Forecast tonight

USSD session / *384*2026#

Tonight

Field Alerts

Back-test / Real collapses, real rainfall

Would it have warned them?

Date	Incident	Rain that day	Rain over 7 days	First rule	Current rule