A lender pays against progress it cannot see
A monitoring surveyor visits periodically and reports on what they saw. Between visits, the position is whatever the borrower says it is.
The problem
Nobody can tell you what a site looked like on Tuesday.
Progress on a housing site is reported by the people building it, in words, once a week or once a month. By the time it reaches a commercial manager or a lender it is a number in a spreadsheet with nothing behind it, and the only way to check any of it is to walk the site.
Progress is self-reportedby the people whose job it is to be on schedule, which is not a criticism of them and is still a reason to want a second source.
A walk is a sample and a daySomebody covers what they can, writes down what they notice, and nothing is recorded that a third party can look at afterwards.
Nobody can show you the evidence for a stageA plot is called "slab" in a report and there is no picture attached to the claim.
Why this gets commissioned
Money on a housing site moves on stage, not on time.
Valuations, drawdowns and subcontractor payments all hang on how far each plot has got. That is a question about a hundred small things spread over a muddy field, and the people who need the answer are usually not the people standing in it.
A commercial manager runs several sites at once
and cannot be on all of them. The reporting they rely on is the thing least likely to carry a picture.
A disagreement about a stage is expensive and late
because it surfaces at valuation, when nobody has a record of what the plot looked like at the time.
A drone pass is cheap and repeatable
which is the whole reason this is worth trying: the same flight next month is a comparison, not a fresh start.
What it changes
What it changes.
Every plot, not a sample
113 of them found and placed from one pass, rather than whatever a walk covered.
A picture behind every stage
Each pin opens the frame that plot was seen best in, so a stage is a claim somebody else can check.
A second source that is not the site team
Read from the footage by software and scored against a person’s labels, with both shown.
A base to fly again against
One pass is one moment. The value is in the second one.
What it is used to decide
The questions it is there to answer.
None of them is settled by the software. Each is a person deciding, with a dated picture of every plot instead of a line in a report.
How far has each plot actually got?Foundations, slab, structure or roofed, per plot, with the frame behind it. On this site: 84 roofed, 9 at structure, 12 at slab and 8 at foundations, counted by eye.
Does the report match the site?Compare the stage counts against what was reported. Where they disagree you have a picture of the plot in question.
What is left, and where is it?29 plots unfinished on this pass, and the model knows where each one is.
Can I get to it?A walkable, caution and blocked map of the ground with a route from the site cabin. Read the limits on this one before relying on it.
The readings
What it found, and how it scored.
Every figure on this page says what kind of figure it is. That is the whole of why it can be used.
113plots found and placed in 3DMeasured, from 3,152 detections collapsed to 120 places. The other 7 were cabins, material stacks, a sign and a mound, and none of the 7 was recognised as not being a plot.
44sof drone footage in, 131 frames, all placedMeasured. The site rebuilt in 5 minutes 48 seconds, about 15 minutes end to end on one graphics card.
81%of plots given the right stage by softwareChecked by a person against labels made by eye. Read that beside the next figure or it means nothing.
70%what you score by calling every plot “roofed”Measured. Three quarters of this site is roofed, so the baseline is high and the 81% is a real but modest gain. On the held-out half it is 77% against 68%.
25 of 29unfinished plots read correctlyChecked by a person. The baseline above gets 0 of 29, which is the part that matters: the unfinished plots are the ones anybody is asking about.
Guides only
And what is only a guide.
These are the readings that are weaker, and they are here rather than quietly left out. Each one is an estimate for an engineer or a site manager to confirm, not a finding.
15 of 19site machines foundEstimate, checked by eye. Vans and material stacks were the misses.
12 of 23material stacks foundEstimate, checked by eye. Trenches and foundations were taken for stacks.
46 of 60squares of ground given the right walkabilityEstimate, and softer than it looks: the eye labels for this one were made after seeing the model’s answers, which is marking your own homework. Against 36 of 60 for calling everything “caution”.
±20%on every distance on the siteEstimate. Scale comes from parked cars of assumed size, not from ground control.
Honestly
What this is not.
Not a replacement for the site team
It is a second source with a picture behind every claim, which is a different thing from a report by somebody who was there.
Not a safety assessment
The walkable and blocked map is a planning sketch. It is not a sign-off and should never be used as one.
Where a surveyed figure exists, it wins
Distances rest on parked cars and carry about twenty per cent. They are for finding things and planning a walk, not for setting anything out.
No saving is claimed, because none was measured
Nothing here was compared against a site walk or a monitoring visit, so there is no figure for what it saves and we are not inventing one.
Built on stock footage, as our own demonstration. Nobody commissioned it: it exists to show what the method does on a working industrial site rather than on a tidy one. The site and its operator are not named here and are not ours to name.
Questions we get asked
Before you ask.
Is 81 per cent any good?
It is a modest gain, and the page prints the number that says so. Calling every plot "roofed" without looking at anything scores 70 per cent on this site, because three quarters of it is roofed. The gain that matters is on the unfinished plots: 25 of 29 right against 0 of 29 for the lazy answer, and those are the plots anybody is actually asking about.
What does it still get wrong?
It calls 12 of the 84 finished houses "structure", which is why the page shows the labels made by eye first and the software’s reading beside them rather than instead of them. It also recognised none of the 7 things that were not plots: a cabin, some material stacks, a sign and a mound all got pins.
What failed completely?
Finding site cabins. It got 3 of 84 right, calling houses and slabs cabins, and we dropped the feature rather than ship it. The first three ways of reading the stage failed too: one answered "slab" 117 times out of 120 and was unusable. It only worked once the question was split in two, asking about the roof first.
Can I trust the walking map?
Treat it as a planning sketch and not a safety sign-off. Its score is also the weakest thing on this page, because the eye labels for it were made after seeing the model’s answers. We are telling you that rather than letting the number stand on its own.
Is this a timelapse?
No. One flight is one moment. The sequence view orders plots by stage, which reads like time and is actually space. A real progress record needs repeat flights, which is the obvious next thing to do and has not been done here.
How accurate are the distances?
About twenty per cent either way. Scale comes from parked cars of assumed size rather than from ground control, so the walk of about 200 metres from the cabin to a typical unfinished plot, against 169 metres in a straight line, carries that margin.
Whose site is it?
Stock footage of a UK-style new-build site. Nobody commissioned this, the site and its builder are not named, and nothing here implies a connection with either.
What would it need from our site?
One drone pass. If you fly your sites already, you have the input. The thing worth doing after that is flying the same site twice.
Got sites you only see once a month?
Tell us how many and how often they are flown. We will say whether this would answer what you need, including when it would not.