AI overview
Stratumly ships AI for the things infrastructure operators actually run: photo defect detection, road and vegetation segmentation on aerial imagery, point-cloud classification, open-vocabulary detection and segmentation, compliance-report narrative, and natural-language queries against your own data.
AI is included in the base subscription, billed as credits consumed per inference. There is no separate AI tier or per-model bolt-on.
Run an AI job
The route in depends on the data:
- Photos: open a survey, go to the Defects tab, click Run defect detection.
- Orthomosaics / GeoTIFF: open the raster, go to the Analysis tab, pick a model (roads, vegetation, spectral index, open-vocabulary).
- Point clouds (LAS / LAZ): open the LiDAR survey, go to the Classification tab, click Classify.
- Free-form prompt against your data: open Ask Stratumly in the top bar and type a natural-language question.
In each case, the system shows the estimated credit cost before you confirm. Run jobs queue and complete asynchronously; you get a notification when results land.
Deliverability tiers
Every AI feature in the product carries one of five badges so you know what kind of model is producing the result:
| Tier | Meaning |
|---|---|
| Live | A trained model running today against your data, with known accuracy metrics published in the model card. |
| Public corpus | A model trained on a public corpus. Works out of the box for the things the public corpus saw; may need fine-tuning for your specific vertical. |
| Needs prep | The pipeline exists but the model is not yet trained for your data. Talk to us about a fine-tune. |
| Customer data | A model trained on accepted / rejected feedback from your organisation. Improves as you use it. |
| LLM-or-rule-based | The result comes from an LLM (Claude) or a deterministic rule, not a trained model. Useful but not statistical. |
Hover the tier badge on any result to see the model card, the corpus, the accuracy metric, and the last training date.
AI credits
One credit ≈ £0.04 internal cost. Typical sizes:
- 10 MP raster processed = 1 credit.
- 2 million LiDAR points processed = 1 credit.
- One LLM-generated report section ≈ 0.5 to 2 credits depending on length.
- One open-vocabulary inference (one image, one prompt) ≈ 1 credit.
Top-up packs are available from Settings → Billing:
- Starter: £10.
- Pro: £20.
- Bulk: £50.
- Power: £200.
Credits never expire. The trial includes a credit allowance so you can run real jobs before paying.
Available models
- Defect detection on photos: YOLOv8n trained on field-defect photos.
- Point-cloud classification: RandLA-Net trained on UK DEFRA LiDAR.
- Open-vocabulary detection and segmentation: SAM2 + GroundingDINO + CLIP on cloud GPU.
- Compliance report narrative: Claude-generated narrative for regulator templates.
Provide feedback on a result
Most AI results have Accept / Reject controls on each detection, segment, or label. Use them:
- Accept marks the prediction as correct in your tenant's feedback log.
- Reject (with an optional reason) marks it as wrong.
This feedback is the input to per-vertical fine-tuning. After enough feedback accumulates in your vertical, your tenant gets its own fine-tuned model that improves on the public-corpus baseline.