Exploration companies sit on large volumes of geological, drilling, and geophysical data, but much of it arrives in inconsistent formats, with missing context, uneven quality, and customer-specific schemas. Before teams can use that information to model mineralisation, predict targets, or plan drilling, they need a reliable way to interpret, validate, and standardise it.
This project defined an AI-assisted workflow that combined schema interpretation, quality scoring, human review, and predictive targeting to help exploration teams move from fragmented raw data to trusted, decision-ready target recommendations. Rather than treating ingestion as a purely technical ETL problem, the work reframed the challenge as a product experience problem: how might we help experts review AI output confidently, resolve ambiguity efficiently, and turn historical and live data into a more actionable basis for the next drilling campaign?
Geological exploration is inherently risky and expensive. Before designing the solution, we needed to understand how geologists currently synthesise disparate datasets to make targeting decisions, and where AI could actually add value without feeling like a “black box” replacing their expertise.
Incoming exploration datasets are rarely clean. Files vary by customer, program, and vendor. Column names differ, units are inconsistent, coordinate systems may be missing, and file types are not always obvious. At the same time, exploration teams need to combine this information with historical and live geophysics to build a clearer picture of where to drill next. Engineering had already begun defining architecture patterns for ingestion, but the broader product challenge was still unresolved:
What I was solving for across the workflow:
Trust in AI-generated interpretation
The system used LLMs to infer file type, schema mapping, and canonical field alignment, but geological professionals could not act on opaque AI output. They needed to understand what the AI thought, how confident it was, where it might be wrong, and what evidence supported a decision.
Human review without excessive friction
If every file required deep manual intervention, the process became a bottleneck. If too much was automated without clear review, poor data could contaminate downstream workflows. The review experience needed to be targeted, explainable, efficient, and increasingly exception-driven over time.
From trusted data to target decisions
The workflow could not stop at verification. The product also needed to show how trusted geological, drilling, and geophysical data could support target prediction and drill campaign planning, with rationale users could inspect and challenge.
Safe downstream consumption
Once data entered a catalogue or visualisation workflow, users needed to know whether it was safe to use. That meant clear confidence, provenance, and verification cues without overwhelming the interface.
The core challenge was designing an interface that surfaces complex AI operations while keeping the user in control. We structured the application around a phased “Exploration Workflow.”
Before AI could generate target recommendations, the system first had to make incoming data usable. Exploration datasets often arrived with inconsistent schemas, unclear file types, missing metadata, and variable quality, which made them difficult to trust or compare.
This phase focused on designing the operational layer that helped teams connect new sources, triage incoming files, assess quality, and identify which datasets were ready for downstream use. Rather than hiding this complexity behind a black box, the workflow surfaced clear signals around progress, confidence, and remediation so users could understand what had been processed, what required intervention, and what could safely move forward.
The goal was to reduce manual wrangling without removing expert judgement. By turning ingestion into a visible, reviewable product experience, this phase created the trust foundation needed for later targeting and planning workflows.
Once data had been reviewed, standardised, and enriched, the workflow shifted from preparation to decision support. This phase explored how trusted drilling, geological, and geophysical inputs could be brought together into AI-assisted targeting scenarios.
Rather than presenting predictions as opaque outputs, the interface was designed to make the recommendation legible: what scenario was being modelled, which signals informed it, and how the resulting targets affected the next drilling campaign. In this case, the system supported copper-focused targeting by combining reviewed datasets with historical and live geophysical context to highlight where confidence was strongest and where planned holes could be validated, deprioritised, or reconsidered.
This was an important design move. The experience was not just about visualising a model, it was about helping exploration teams connect evidence, confidence, and action in a way that felt credible enough to support real planning decisions.
This work helped turn a backend-heavy discussion about ingestion and AI into a clearer product model that connected data quality, human review, and predictive targeting in one coherent workflow.
Instead of treating ingestion as a hidden preprocessing layer, the prototype showed how review, validation, and trust signals could be designed as first-class product experiences. It also demonstrated how those upstream decisions could flow into more actionable targeting outputs, helping exploration teams understand not just where data came from, but how it could support the next drilling campaign.
At a product level, the prototype established a stronger story for Exosphere: AI was not positioned as a vague assistant layered on top of geological workflows, but as part of an operational system that accelerated interpretation, preserved human judgement, and made downstream targeting more usable and explainable.