Construction Progress Tracking with Computer Vision

It depends on the trade and how visible the work is. Structural stages — excavation, foundations, framing, cladding — read reliably because they change the shape of the building. Finishing work behind a wall does not, and no camera will tell you otherwise. We measure accuracy against your own current process on a sample of your photos before you commit to anything, so you get a real number for your sites rather than ours.

No. The system is built around the photographs your team already takes on their phones. Fixed cameras give you more frequent coverage if you want it, and we can work with those too, but they are not a requirement to start.

Those are the normal case on a building site, not the exception, so the model is trained on them. What matters more is that anything it is unsure about is flagged rather than guessed — a low-confidence reading goes to a person instead of quietly entering the report as fact.

Between neighbouring stages, usually. Between stages that look identical from outside, no — and we would rather say so. Where the visual difference is genuinely absent, the honest answer is that the photo cannot carry the information, and we will tell you that during the assessment rather than after you have bought something.

The output is structured data, so it goes wherever your reporting already lives. We push into Procore and Autodesk Construction Cloud, write to a spreadsheet, or hand you a REST endpoint and let your team wire it up. We do not ask you to move off the tools you run.

Both depend on how many sites you run and how much of the reporting you want automated. Every project starts with a free assessment: send us a sample of your site photos and we will tell you what can be read from them, how accurately, and what the reporting time saved is worth. You will get that back within 24 hours, from the engineer who would build it.

Both, but they want different things. Site teams want the exceptions: what slipped, what is blocked, what needs a hard hat conversation this morning. Head office wants the series: the same measure, week after week, across every job, so a slipping programme is visible before it becomes a claim. Because the output is structured data rather than prose, the same run feeds both without anyone re-keying it.

PPE scoring needs to see people, so faces appear in the source images. We treat that as a data protection question rather than a technical detail: the images stay in your tenancy, the model returns counts and compliance flags rather than identities, and we can blur faces on ingest where your works council or GDPR assessment requires it. That gets agreed before any photographs move, not after.

That is the normal way in, and we would push back on anyone who wanted to start wider. One job, a few weeks of photographs you have already taken, measured against the reports your team wrote by hand for the same period. If the numbers do not hold up on your own sites, you have lost a fortnight and no money.