Field Non-Compliance Analysis

The subcontractor’s report, scanned as a PDF

On a construction site, a non-compliance rarely arrives in a clean form. It shows up as a phone photo, a paper form photographed at arm’s length, or three rushed lines in an end-of-day report, sometimes written by a temporary worker unfamiliar with internal codes. The HSE officer recompiles these scattered sources into a spreadsheet every week before the site meeting, retyping by hand whatever isn’t already in a usable format.

The pilot works well on the tidy reports

A demo always runs on the site’s best reports: the ones that follow the template, written up by a diligent site engineer. In production, the same model meets forms filled in pencil, abbreviations specific to one subcontractor, photos with no caption. Classification quality drops, and teams fall back into reading everything themselves. Beyond the wasted effort, a serious non-compliance that gets misfiled, buried among duplicates or false positives, can sit unaddressed longer than it would have under the original spreadsheet.

One simple condition, often overlooked

The project is worth building only if a minimum set of fields is consistently filled in: location, date, nature of the deviation, even as free text. Without that baseline, the model guesses more than it classifies. It also needs a named quality or HSE owner to settle ambiguous cases, plus prior agreement on the severity typologies used, or the categories drift within weeks and every site ends up inventing its own.

From flag to escalation decision

We build a pipeline that ingests reports and photos, extracts the type of deviation, the area, and a severity estimate, then groups recurrences by site or by work package. The HSE officer receives a list ranked by severity and area, and decides alone which non-compliances go up to the project steering committee. Cases the model cannot classify with enough confidence are flagged as such, rather than defaulted into a minor category.