Snapshot
Who: Midwest grain cooperative (anonymized) — agronomy desk supporting member farms.
Problem: Imagery piled up on drives; scouts couldn’t cover priority fields first; variable-rate prescriptions lagged the weather window.
What we shipped: Automated ortho stitching pipeline, NDVI + thermal fusion models, and a mobile triage list ranked by expected margin impact.
Challenge
Trust was fragile: a bad recommendation in one county could lose the whole zip code. Models had to show provenance and confidence.
Approach
We co-designed “safe modes” with the chief agronomist—no auto-spray hooks until two seasons of shadow recommendations matched scout notes.
Solution components
Rollout timeline
Season 1 · Shadow
Model outputs visible only to desk staff; weekly calibration vs. scout routes.
Season 2 · Pilot counties
Push alerts for top 10% of at-risk acres; spray decisions still human-owned.
Season 3 · Scale
Integrated prescription export to two leading FMIS vendors.
“We finally had one screen that told us where to drive first on Monday—not fifty PDFs from three vendors.”
— Lead Agronomist, anonymized cooperative
Results & controls
- Pilot acres +12% yield vs. matched control blocks (same hybrid cohort).
- Fungicide spend −37% where pressure models stayed below action thresholds.
- Median time from anomaly to first scout visit dropped from 9 days to 2.
- Every alert stores source layers (NDVI frame IDs, probe IDs, grid cell).
Representative scenario for marketing purposes. Agronomic outcomes vary materially by region, hybrid, and season.