We treat permits as ground truth - then verify.

A language model classifies roof-related permits to distinguish full reroofs from partial work or unrelated projects. We start at the original construction year, sequence permits forward to identify true replacements, and cross-check against imagery to avoid stale or incomplete records. This flags fraud, catches data-entry errors, and grounds predictions in current roof conditions.

How it works:

• Classify: Separate full replacements from patches, solar, or additions.

• Sequence: Walk forward from construction year to the latest verified replacement.

• Cross-check: Confirm completion in imagery and surface unpermitted or incomplete jobs.

Outcome: Verified replacement dates for higher-precision roof age.

Computer vision confirms true reroofs across 20+ years of imagery.

Models trained on thousands of human-labeled events detect full material changes while ignoring lookalikes like patches, solar installs, and additions. A strict QC pipeline drops blurry or outdated images and anchors on consistent visual signals, delivering deeper coverage—including rural areas—and far fewer false positives.

How it works:

• Long horizon: Identify earliest clear visual evidence to set minimum roof age.

• Full vs. partial: Detect true reroofs and filter out non-replacement activity.

• QC pipeline: Image-quality checks ensure only reliable imagery feeds the model.

Outcome: Broad, reliable confirmation at national scale (~95% coverage, ~92% accuracy).

If no permit exists and imagery shows no material change, we estimate.

We establish a minimum-age floor from the earliest clear visual evidence, then model realistic aging paths using the latest imagery, observed roof condition, property context, and regional climatology (UV, humidity, freeze–thaw cycles, wind and hail exposure).

A roof in the Southeast ages differently from one in the Pacific Northwest—and the model reflects that.

How it works:

• Trigger: No permit + no visible material change.

• Inputs: Minimum roof age, latest imagery, roof condition, property context, and regional climate factors.

• Regional climatology: Locale-specific weathering patterns tighten age estimates.

Outcome: A calibrated, defensible roof-age estimate—even when direct signals are absent.

Every prediction includes a confidence score to drive risk-based workflows.

High-confidence results can move straight through quoting or rating; medium-confidence results trigger lightweight review; low-confidence results route to inspection. You define how each confidence band flows through your underwriting process.

Banding

High → auto; Medium → secondary review; Low → inspection.

Governance

Transparent thresholds for underwriting and compliance.

Outcome

Risk-based workflow, less friction, better allocation of inspections.