A consultant's assessment takes 7 to 15 working days. The decision it feeds is made in hours. This closes that gap — and it does it without a generative model anywhere near the drawing.
Feasibility studies, land acquisitions and bid submissions all need early plant room sizing.
Those spaces drive building configuration and saleable floor area — they are not a detail.
A consultant assessment takes 7 to 15 working days.
Internal design managers are committed to live projects and cannot turn one around.
So the estimate that shapes the scheme gets made without engineering input, or the bid waits.
Each revision costs billable hours, so options get rationed and the first workable layout wins.
Strategies propose placements. A geometry engine validates them. Nothing is invented.
The catalogue is built on published manufacturer footprints and service clearances, including the tube-pull zone off one short end — which is what actually drives spacing in a plant room. A headroom check runs against unit height plus overhead clearance.
Deterministic packing strategies suggest positions — row-major, column-major, rotated. A geometry engine then checks every candidate against the room envelope and against every other unit. The governing rule is the subtle one: a unit body may not enter another unit's keep-out, but two keep-outs may legitimately overlap, because a service aisle is shared.
Valid layouts are scored on space utilisation and compactness, so the winner leaves contiguous free floor for expansion rather than scattered gaps. The engine returns the best layout plus alternatives, as a scaled floor plan and as structured data. An infeasible brief returns a reason, not a silent failure.
The obvious build was a generative one: describe the room, let an image model produce the layout. It worked well enough to demo. It was replaced with about 800 lines of deterministic geometry, for two reasons — and the second matters more than the first.
Cost per revision. The entire value proposition is that you can explore options freely: re-run the same brief under a different fire code, a different redundancy standard, tighter clearances, lower headroom. That only works if a revision is nearly free. Generative image calls are not nearly free, and the economics collapse exactly where the product is supposed to be strongest.
Verifiability. A generated image can only be believed. A drawing derived from exact published dimensions can be checked — unit counts, clearances and service gaps are correct by construction, and an engineer can measure them straight off the output. For work that feeds a real investment decision, that difference is the whole thing.
"The problem was geometry, not imagery. The obvious tool was the wrong tool."
Where AI does belong here is the input edge, not the output: reading a brief in plain language, pulling constraints out of a standards document, answering questions about why a layout came out the way it did. That is the next phase. The geometry stays deterministic.
| Metric | Before | After |
|---|---|---|
| Time to a concept layout | 7–15 working days | Minutes |
| Cost per revision | Billable hours or queue time | Near zero |
| Options explored | Rationed — first workable layout wins | As many as the site deserves |
| Changing the rule set | A new engagement | Re-run under a different code or redundancy standard |
| Design review | "We'll revise and come back" | Regenerated in the meeting |
| Consistency between projects | Varies by person and by month | Same rules applied every time, checkable against the drawing |
Scope: output is conceptual, for planning, feasibility and bid stages. Detailed engineering, statutory approval and stamped drawings remain with a licensed engineer. The tool is designed to make the early decision better informed, not to replace the person who signs the drawing. Currently in a funded pilot across three asset classes.
"If the output has to be checked, it has to be constructed — not generated."
There is a strong pull right now to put a model at the centre of every system. Sometimes that is right. Here it was not: the work was geometric, the constraints were known and published, and the answer needed to be verifiable by someone who would be held responsible for it.
The useful test is simple. Ask what happens if the output is subtly wrong and nobody notices. If the answer is a bad afternoon, generation is fine. If the answer is a building that does not work, the calculation should be deterministic and AI should sit at the edges — reading the brief, explaining the result, handling the language — where being approximately right is genuinely good enough.
Racking, aisle widths, turning circles and pick paths are all hard constraints with published dimensions. The same propose-validate-score loop applies directly.
Treatment rooms, circulation, accessibility clearances and equipment envelopes, tested against several room shapes before a lease is signed.
Any plant space where service access, maintenance withdrawal and statutory clearance drive the footprint more than the equipment itself does.
Testing many candidate sites quickly against the same rule set, to find which ones fail on technical space before anyone spends money on them.
90-minute AI Opportunity Audit. No pitch — just findings you can act on.