Engagement snapshot
- Mandate
- Speed up early scoping without letting image interpretation become an unsafe automation shortcut.
- Timeline
- 13 weeks from data prep to three-region deployment.
- Team shape
- Operations VP, 1 computer vision engineer, 2 full-stack engineers, and 6 field reviewers.
The problem
Estimators waited on senior reviewers to interpret water and fire damage photos before deciding what crews, materials, and inspection steps should go first.
What we built
Deployed a multimodal triage layer that tagged visible damage patterns, proposed scope starters, and flagged image sets that needed human review before dispatch.
Operating context
The company served a large geographic footprint with variable field staff experience. Early scoping delays created slower dispatch and expensive second visits when the first crew arrived underprepared.
Key constraints
- Photo quality varied heavily across contractors, lighting conditions, and mobile devices.
- No automated recommendation could skip mandatory moisture, safety, or hazard inspection steps.
- The system had to explain what it saw well enough for estimators to trust or reject it quickly.
What we built
Multimodal evidence parsing
Combined image interpretation with field-note text and job metadata before proposing scope hints.
Category-specific recommendations
Separated water, smoke, mold, and structural cues so scope suggestions stayed tied to the actual damage class.
Reviewer-in-the-loop controls
Required signoff on low-confidence photo sets and surfaced visible rationale for every suggested next step.
Delivery path
Training data curation
Selected representative job photo sets across damage types, lighting conditions, and device quality to avoid brittle performance.
Estimator workflow fit
Designed outputs around the estimator intake screen rather than a separate AI dashboard.
Dispatch feedback loop
Tracked whether crew dispatch recommendations reduced second visits and scope corrections after deployment.
Why it mattered
The system reduced the time to a credible first scope and helped standardize work between experienced and less experienced teams. The biggest operational value came from better prepared first dispatches.
Implementation notes
- Multimodal systems need explicit fallback behavior for poor image quality, not just better models.
- Visual explanations and tagged evidence materially improved estimator trust.
- Scoping speed only mattered because the dispatch workflow consumed the recommendations immediately.