Illustrative demonstration · Synthetic dataOrganisations, transactions, documents, figures and outcomes shown here are fictional or modelled for demonstration purposes. No client, bank or confidential data is used. Results are illustrative, not actual client outcomes or professional advice.
Industry case · Mining & resources · Interactive demo
From flight hours to pit-floor P&L.
A 12-month drone programme at one open-pit iron-ore site: stockpile surveys, pit-wall monitoring and haul-road inspection, read as one dataset and then converted into five costed moves. The data is the diagnosis; the strategy is the treatment.
Headline insightTwo haul-road segments, H3 and H5, produce 61% of tyre-damage incidents and sit in the top defect-density band all four quarters. Regrading just those 3.1 km is the single cheapest lever on the site: ~$470k a year in tyre and fuel effect for ~$85k of works.
1Read the operation2Price the moves3The strategy
1 · Read the operationWhat the flight data actually says
Five views, one dataset, and each ends in a move
Survey coverage and reconciliation error, month by month
Hectares surveyed (bars) against stockpile volume error vs ground truth (line)
Hectares surveyedVolume error %
Insight. The error step-change in May tracks the move from monthly to fortnightly flights, not better hardware. → Cadence, not sensors, is the cheapest accuracy lever; extend fortnightly cadence to the other two pits (Move 1).
Stockpile volumes: drone vs manual survey
End-of-quarter volumes for the four largest ROM stockpiles, kt
Manual surveyDrone photogrammetry
Insight. Manual surveys over-state Stockpile C by ~5%: a stale compaction factor, ~25 kt of phantom inventory on the books. → Fix the density assumption before quarter-end to avoid a write-down surprise (Move 4).
Haul-road defects: Pareto view
Computer-vision detections by defect class (bars) with cumulative share (line)
DetectionsCumulative %
Insight. Rutting and edge erosion alone are 58% of all defects: a grading and drainage problem, not a surface-material problem. → Shift budget from resheeting to grader hours and drainage on the worst segments (Move 2).
Defect density by haul segment and quarter
Defects per km per quarter, detections mapped to six haul segments
LowHigh
Insight. H3 and H5 stay in the top band every quarter: chronic, not seasonal. Cross-referenced with dispatch data, they carry the loaded downhill runs. → This heatmap is the works order: it converts "roads are bad" into "fix these 3.1 km first".
Pit-wall movement alerts and response time
Weekly change-detection alerts (bars) and mean time from alert to geotech triage, hours (line)
Movement alerts / weekMean triage time, hours
Insight. Alert spikes follow blasting weeks, as expected. The value is in the triage line: auto-routing alerts with imagery attached cut response from ~31 h to ~9 h. → Faster triage means benches re-open sooner after precautionary closures: ~$260k a year of avoided standby (Move 3).
2 · Price the movesWhere should effort go first?
Toggle moves in Section 3; the matrix, the bridge and the totals respond
Effort vs modelled annual value
Bubble size = confidence in the estimate · dashed = risk removal, no direct value modelled
Insight. Two initiatives sit in the "do now" corner: haul-segment regrading and alert auto-routing. Hyperspectral dilution mapping is the biggest prize but the hardest build: pilot it, don't bet on it.
Modelled annual value bridge
How the selected moves stack, USD thousand per year · lighter bar = pilot upside
3 · The strategyFive moves, sequenced cheapest-certain first
Click a move to include or exclude it from the programme
Guardrail · GovernanceEvery model output gets a human ownerDefect detections, movement alerts and volume models inform decisions; they don't make them. Each feed has a named reviewer, an accuracy-check cadence and a documented override path. How: a monthly model-health review of detection precision, alert false-positive rate and survey-vs-ground-truth drift.
