What an AI Prospectivity Engagement Actually Looks Like

The most common question after the first three articles in this series is also the most practical. On the client side, what does it actually take to run a DepoDart project, 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 usually eat up the schedule.

A pilot is a short, hands-on project. The client shares a dataset for an area of interest. Our team works alongside theirs and handles the data loading, the clean-up, the modelling, and the review. At the end, the client leaves with a prospectivity map, ranked drill targets, uncertainty bounds, and a full record of where each result came from. A typical project runs two to four weeks, depending on how much data there is, how complex it is, and how much back-and-forth the review with your team needs.

Data sharing

The client sends whatever is in the archive: any format, any age, any coordinate system. If you need an NDA before the transfer, one is available. You do not need to clean up, reproject, or reformat anything. Our team is built to take in the full mix of a real exploration archive, including geophysics, soil chemistry, drill logs, geological maps, and satellite data.

The bar for this step is low on purpose. A team that spends two weeks tidying data before sending it is doing work we are meant to do for them.

Ingestion and normalisation

Our team loads the dataset, lines it up across formats and coordinate systems, and brings every layer into one shared spatial model. The client gets a data receipt that confirms what was loaded, at what resolution, and over what area. This is the moment when the hidden model inside the archive becomes a clear one.

If anything in the dataset cannot be read, is unclear, or needs a question answered, it shows up at this stage, while there is still time to sort it out together before the model run.

AI model run

The combined dataset is scored by a group of models called gradient-boosted decision trees, trained with positive-unlabelled learning. This is the standard approach when known deposits are very rare. The scores are then calibrated with Platt scaling. Testing uses separate map blocks rather than random points, so the confidence we report reflects what the model will really do on new ground.

Every point in the area gets a ranked probability score for mineralisation. Regional datasets come out at 50 to 100 metre pixels. High-resolution private datasets come out at 10 to 25 metre pixels. The result is a smooth prospectivity map, at district scale and finer.

QC and geoscientist review

Before anything is delivered, a qualified geoscientist on our side reviews the output, often in conversation with the client's geologists. The review checks that the model makes sense against the local geology, that the top targets are not artefacts, and that the uncertainty layer matches where the data actually covers. Anything that looks wrong is flagged in the delivery, not hidden in it.

This step exists because the work is built by geologists for geologists. A model result that no geoscientist has read before it ships is not a deliverable. It is a draft.

Delivery

Every project ships four deliverables. The first is a prospectivity map as a GeoTIFF, a colour-coded image with the chance of mineralisation in each pixel. The second is a ranked drill target list as a PDF and CSV, with probability scores, mineral type, estimated depth, and a confidence tier. The third is an uncertainty map, a matching GeoTIFF that measures the model's uncertainty in each pixel. The fourth is a data provenance report, a full trail showing which input layers drove each recommendation.

All four are made to drop straight 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 a single target in front of a technical committee, not just show a number.

What the client keeps

At the end of the project, the deliverables belong to the client. The data you sent is yours, the maps that came back are yours, and the targets are yours. There is no software to install and no login to maintain. The pilot is built to answer one narrow question: whether the work is worth building into your next program.

How we work today

Right now we work closely with each client, hands on. We are small on purpose, and that lets us shape every project around what your team actually needs, rather than around a fixed process. Over time, more of this will become self-serve. For now, the work that earns your trust is built to be done together.

What this means for your next program

A pilot fits in the gap between board approval and the start of a field season. The deliverables are solid enough to judge on their own. A team that starts a project in the spring has a combined model, a ranked target list, and an uncertainty map in hand before the rig even moves.