Snapshot
Who: Open-pit iron operation (anonymized) — mixed fleet, two shovels feeding three cells.
Problem: Grade control models refreshed too slowly; shovel operators chased visual color that didn’t match the block model.
What we shipped: Nightly block model fusion, uncertainty surfaces for planners, and a tablet view that translated geostatistics into simple “drill here first” bands.
Challenge
Mine engineers were skeptical of black-box ML—anything that couldn’t show variograms and search parameters was dead on arrival.
Approach
We wrapped models in the same QC gates their consultants already used, and co-signed the first six patterns with the external geology firm.
Solution components
Rollout timeline
Weeks 1–3 · Data contracts
Normalized collar tables, fixed CRS drift between survey vendors.
Weeks 4–8 · Shadow patterns
Printed maps for bench meetings; no blast changes yet.
Weeks 9–12 · Controlled blasts
A/B sectors with post-blast muckpile sampling.
Quarter 2 · Operations embed
INSART engineer on-site two days/week during wet season.
“The block model finally met the muckpile. We still blast—but we aim with both eyes open now.”
— Chief Geologist, anonymized operation
Results & controls
- 31% more mass inside the ore envelope for the same drilled meters (bench trial cohort).
- Waste tons to stockpile −18%; crusher feed variability down measurably week-over-week.
- Haul fleet fuel −7% after routing avoided rework loops on wet days.
- Every pattern stored with variography parameters and approver signatures.
Representative scenario for marketing purposes. Mining economics and permitting contexts vary widely by jurisdiction.