The photographs you already have
Field service, construction, installation and maintenance teams routinely photograph completed work. Those images are stored against the job and almost never looked at again unless there is a dispute.
They are a substantial dataset. The realistic question is not whether a model can judge quality, but which specific, well-defined checks can be automated reliably.
What is reliably checkable
| Check | Feasibility |
|---|---|
| Is there a photo at all, and is it usable? | High - blur and darkness detection is easy |
| Does it show the expected type of equipment? | High with training data |
| Is a required element present - label, seal, guard? | Good |
| Does the location metadata match the job site? | High - metadata, not vision |
| Has this photo been submitted before? | High - duplicate detection |
| Does the workmanship meet standard? | Low - needs an inspector |
The first and last rows are the important contrast. Confirming a usable photo exists showing the right equipment removes a large share of administrative checking. Judging quality is a different proposition.
The unglamorous checks pay first
Blur, darkness, a photo of the van floor, the same image submitted for three jobs, missing metadata - these are common and easy to detect, and catching them at submission time is worth more than any sophisticated analysis.
The reason is timing. A problem caught while the engineer is still on site is fixed in seconds. The same problem found in an audit three weeks later means a return visit.
Practical collection problems
- Lighting varies enormously between a plant room and an outdoor installation
- Angle and distance are inconsistent without guidance
- Phone cameras differ across a fleet of devices
- Connectivity on site delays or loses uploads
- Privacy - people and third-party property appear in the background
Several of these are solved by guidance rather than modelling. An overlay in the app showing the expected framing improves consistency more than any amount of training data, and makes every downstream check easier.
Support the inspector, do not replace them
For anything approaching a quality judgement, the sensible design prioritises which jobs a human inspects rather than passing or failing work automatically.
That keeps accountability with a person, which matters where safety or compliance is involved, while still concentrating inspection effort where it is most likely to find something. It also means model errors cost inspection time rather than creating a wrongly approved installation.
Catch the unusable photo while the engineer is still on site. Everything after that costs a return visit.