Project case study
Inspection Intelligence
Manual inspection time dropped from 45+ minutes to under 3 minutes. End-to-end system: video frame extraction, damage detection, and LLM-generated reports with mandatory human review.
Private implementation; details are shared at an architectural level.
Problem
- Inspectors were spending 45+ minutes per walkthrough on photo comparison and manual report writing
- Across thousands of walkthroughs, this was the single biggest bottleneck in the downstream workflow
Constraints
- Has to work with existing mobile capture, no new hardware
- Target was under 5 minutes per walkthrough
- Video quality and lighting vary widely
- Zero tolerance for missed damage in a workflow where the output drives real financial decisions
Architecture
- Backend service extracts frames from walkthrough videos at configurable intervals
- Managed vision models handle object detection and image comparison against a baseline
- LLM layer turns detected anomalies into a structured damage report
- Review UI gives inspectors a side-by-side view with one-click override on any item
Tradeoffs
- Frame extraction over real-time video analysis. Simpler, roughly 10x cheaper, accurate enough for the use case
- Managed vision models over custom-trained, to avoid a training-data maintenance burden for V1
- Mandatory human review over full automation. The 5% of edge cases are the ones that actually cost money
Outcome
- ~95% reduction in manual inspection time
- Under 3 minutes per walkthrough
- Human review catches edge cases without slowing the pipeline