Prepare tiles
Pin imagery tiles and labels; hold out labelled regions for evaluation.
Industry blueprint
Geospatial and
remote sensing
Agriculture, forestry and mining. Fine-tune foundation models on scarce labels and prove the gain.
Labelled imagery is rare, so teams fine-tune pretrained segmentation models on small labelled sets. This design measures before and after fine-tuning so the gain is shown, not assumed.
Illustrative pipeline
Measure the baseline
Score the pretrained model on the held-out regions first.
Fine-tune
Run the partner's segmentation fine-tune on your GPUs inside the budget.
Compare and approve
Promote only if the fine-tuned model beats the baseline and an analyst approves.
What the record keeps
The imagery version and labels, baseline scores, the fine-tuned candidate, after scores, the spend, and the approver.
Partner brings
- Imagery access and labelled tiles
- Segmentation model
- GIS tooling
Cairn brings
- Pinned imagery and held-out regions
- Baseline-then-compare, as in verifiable-finetune
- GPU budget and approval
- A record of every published model
Business value to test
Segmentation accuracy on held-out regions, label cost per gain and compute per run.
Keep the decision explicit
A model is published only after a recorded approval.
Close the feedback loop
Analyst corrections become labels for the next fine-tune.
Reference design, not a shipped domain pack. The measure-train-measure pattern comes from training/verifiable-finetune.
See how failure can trigger a reviewed improvement cycle →Prove the fine-tune helped.
Talk about your imagery, labels and what accuracy you need.