What an AI Prospectivity Engagement Actually Looks Like

The most common question after the first three articles in this series is the most practical one. What does it take, on the client side, to actually run a DepoDart engagement, and what comes back at the end. The answer is shorter than most exploration teams expect, because our team takes on the parts of the work that traditionally consume the timeline.

A pilot is a focused, hands-on engagement. The client shares a dataset covering an area of interest. Our team works alongside theirs, takes on ingestion, normalisation, modelling, and review, and at the end the client walks away with a prospectivity surface, ranked drill targets, uncertainty bounds, and full data provenance. A typical engagement runs two to four weeks, depending on dataset volume, complexity, and how iterative the review with your team needs to be.

Data sharing

The client sends whatever is in the archive. Any format, any vintage, any coordinate system. An NDA is available before transfer for teams who need one in place. The team does not need to preprocess, reproject, or reformat anything. Our team is set up to absorb the heterogeneity of a real exploration archive, including geophysics, geochemistry, drill logs, geological maps, and remote sensing.

The bar for this step is intentionally low. A team that spends two weeks getting data into a clean format before sending it is doing work we are supposed to do for them.

Ingestion and normalisation

Our team ingests the dataset, normalises it across formats and coordinate systems, and resolves every layer into a unified spatial model. The client receives a data receipt confirming what was ingested, at what resolution, and across what footprint. This is the moment the implicit model inside the archive becomes an explicit one.

If anything in the dataset is unreadable, ambiguous, or in a format that needs clarification, this is the stage where it surfaces, while there is still time to resolve it together before the model run.

AI model run

The fused dataset is scored by an ensemble of gradient-boosted decision trees trained with positive-unlabelled learning, the standard approach for spatial classification problems with severe class imbalance. Predictions are calibrated with Platt scaling. Validation uses spatial cross-validation with geographic blocks, not random holdout, so the reported confidence reflects what the model will actually do on unseen ground.

Every point in the area of interest receives a ranked mineralisation probability score. Regional datasets resolve to 50 to 100 metre pixels. High-resolution proprietary datasets resolve to 10 to 25 metre pixels. The output is a continuous prospectivity surface at district scale and below.

QC and geoscientist review

Before anything is delivered, the output is reviewed by a qualified geoscientist on our side, often in dialogue with the client's geologists. The review checks that the model has behaved sensibly against the geology of the area, that the highest-ranked targets are not artefacts, and that the uncertainty layer is consistent with the data coverage. Anything that looks anomalous in a bad way is flagged in the delivery, not hidden in it.

This step exists because the work is built by geologists for geologists. A model output that has not been read by a geoscientist before it ships is not a deliverable, it is a draft.

Delivery

Every engagement ships four artifacts. The first is a prospectivity surface as a GeoTIFF, a colour-coded raster with ore-forming probability per pixel. The second is a ranked drill target list as a PDF and CSV, including probability scores, mineral type, estimated depth, and confidence tier. The third is an uncertainty map as a companion GeoTIFF, quantifying model uncertainty per pixel. The fourth is a data provenance report, a full audit trail showing which input layers drove each recommendation.

All four are designed to drop into the tools an exploration team already uses. The GeoTIFFs open in any GIS. The CSV opens in any database or spreadsheet. The provenance report is written so a geologist can defend an individual target in front of a technical committee, not just present a number.

What the client keeps

At the end of the engagement, the artifacts belong to the client. The data sent is theirs, the maps that came back are theirs, and the targets are theirs. There is no software to install and no platform login to maintain. The pilot is designed to answer one narrow question, which is whether the work is worth incorporating into the next program.

How we work today

At this stage we work closely with each client, hands on. We are deliberately small, and that lets us shape every engagement around what your team actually needs rather than around a fixed workflow. Over time, more of this becomes self-serve. The work that earns your trust now is collaborative by design.

What this means for your next program

A pilot fits inside the window between board approval and the start of a field season. The artifacts are concrete enough to evaluate on their own terms. A team that starts an engagement in the spring has a fused model, a ranked target list, and an uncertainty map in hand before the rig is mobilised.