Building with Confidence: How Technology Changes the Client Experience

Building With Confidence

For years, one constant has defined the construction industry: uncertainty. Whether it is concern about unexpected costs, communication gaps, design changes, project delays, or simply not knowing what the finished building will look like until construction is underway, clients have traditionally carried much of the risk.

Digital tools are changing that, and the change is less glamorous than it sounds. The international standard that governs how project information is handled, ISO 19650-1, describes the goal in almost administrative language: a framework for managing information across the whole life cycle of a built asset, covering how information is organized, exchanged, versioned and recorded by everyone involved.¹ While conversations about innovation tend to center on artificial intelligence, the technologies that change a client's experience most work quietly, by making sure everyone is looking at the same, current, correct information.

Yet technology alone is not what clients are investing in. Clients do not choose a design and build partner because it owns the newest software, the most advanced drone, or the latest laser scanner. They choose a partner because they want confidence that their project will be managed thoughtfully, communicated clearly, and delivered with as few surprises as possible. Technology becomes valuable when it delivers what clients are actually looking for: transparency, coordination, informed decisions, and trust. The competitive advantage sits in integrating the technology into a single process that improves the client's experience from concept to completion. Owning the tools is the easy half.

What follows is written from our own operating experience. WALLNUT runs its own reality capture pipeline and builds its own model to cost software, so what is described here is what we operate, including the parts that do not work yet.

Building Information Modeling (BIM)

BIM is often described as a 3D modeling tool, but that definition misses the point. ISO 19650-1 sets out BIM as a set of business processes for producing and managing information across a built asset's life cycle, from strategic planning and design through construction, operation, maintenance and end of life.¹ The model is the artifact. The management of information is the discipline.

Two things follow from that, and both matter commercially to a client.

The first is the format. IFC, standardized as ISO 16739-1, is the open schema for exchanging building information between the different software applications used across a project.² Working in an open format means a client's model is not hostage to one supplier's file type. Our own viewer is built on open IFC for exactly that reason: the geometry, the properties and the classifications remain readable by anyone the client later chooses to work with.

The second is that a model becomes commercially useful only when it is connected to something. In our platform, elements in the IFC model are linked to the articles of the bill of quantities, so that a line in a measurement certificate and a wall in the model are known to be the same thing. That link is built: our engine resolves an IFC element to the article of the bill of quantities that prices it. What is not built is the screen. Today the viewer colours a model by discipline and by source model, not by cost and not by the percentage of each article executed and certified — that view exists as a library with no interface in front of it, and we are not going to describe it as though you could open it. When it ships, it is the difference between a monthly report that asserts progress and a model that shows it.

The failure modes here are not obvious from a demonstration, so they are worth stating. A quantity read from a model is only as reliable as the modeling convention behind it: if elements are not classified consistently, extraction produces numbers with decimal places and no authority. And models change. Element identifiers can be regenerated when a file is re-saved, so any link between a model and money has to be pinned to a specific model version, or it will quietly start pricing the wrong revision. Building this properly is mostly bookkeeping discipline, not 3D graphics.

Reality Capture

One of the greatest sources of uncertainty in construction, particularly in renovation and refurbishment, is the existing building itself. Old drawings may be wrong, conditions change over decades, and discrepancies often surface only once demolition has begun. This is where reality capture earns its place: it produces a dated digital record of what is actually there before design work begins, and at WALLNUT it is the first thing we do on an existing building.

In practice, we walk the building with a 360 degree camera, select the sharpest frames from the recording, and reconstruct the camera positions photogrammetrically. From that single alignment we derive two different products for two different purposes: a point cloud and a textured mesh, which carry the geometry, and a Gaussian splat, which is what the client actually walks through on screen. The deliverables that leave that process are open ones: E57 point clouds, meshes, and sections and plans exported to CAD.³

What it costs in effort

Reality capture is often presented as pressing record. It is not.

Capture itself is a disciplined procedure. Indoors we fix a fast shutter speed, typically 1/500, because motion blur destroys the reconstruction; outdoors the camera can be left on automatic.⁴ The operator walks slowly and steadily, keeps at least 30 centimeters from any surface, runs a perimeter loop first, fills the interior in a grid, and returns to the starting point so the loop closes, since loop closure is what holds the reconstruction rigid.⁴ Scenes are kept to five to ten minutes each, and a scene typically yields 250 to 1,000 selected frames, roughly one sharp frame per half second of video, with 70 to 80 percent overlap between them.³ Battery life is a real constraint: about 88 minutes of 8K recording per charge.³

Processing is heavier than capture. Alignment of the frames is the slow, decisive step, and in our stack it still runs in a desktop application that our license tier cannot script, so it remains a manual operation at a workstation.⁵ Training a single scene on a dedicated GPU workstation runs from roughly half an hour to a couple of hours depending on frame count and resolution, before any cleanup.⁶ Then someone has to look at the result and delete the artifacts. In our experience a morning of walking a building becomes a full day of processing and checking. That is the honest ratio, and it is why we do not offer capture as a free extra on every visit.

There is also the operator problem, which no vendor brochure mentions. A 360 camera sees the person carrying it in every single frame. The person has to be masked out of the imagery automatically before alignment and again before training, and getting that masking right is a genuine engineering task in its own right.⁷

Where it falls short

Scale does not come for free. A reconstruction built from images alone has no inherent real world size. Scale and position have to be introduced from outside the photographs, either by placing a printed calibration target or a known distance in the capture,⁴ or by registering the reconstruction against known coordinates afterwards.⁸ If nobody does that, the result looks perfect and measures nothing. Any firm that offers to measure from a capture should be able to say which of those two things it did.

Appearance is not geometry. Gaussian splats are exceptionally good at looking like the building. They are not a survey. In a peer reviewed comparison presented at the ISPRS Geospatial Week 2025, classical multi-view stereo reconstruction achieved 1.43 mm RMSE against a ground truth mesh on a controlled indoor object, while the Gaussian splatting methods tested reached 4.74 mm, 5.80 mm and 7.28 mm on the same scene; outdoors, with vegetation in the way, the same methods degraded from 3.23 mm for multi-view stereo to between 7.49 mm and 17.36 mm.⁹ AEC Magazine put the industry version of the same point well: when vendors describe a splat as survey grade, they are describing a workflow where LiDAR or SLAM did the geometric work, and "splats inherit that accuracy, they do not generate it."¹⁰ Our practice follows from this. We measure on the photogrammetric mesh, the point cloud or the BIM model. We do not measure on the splat, and we say so.

It fails, and the failures are invisible to the client. When we audited nine processing runs from our own pipeline in mid 2026, two had ended in outright crashes, one of them after eighteen hours of computation; several more were left stranded because a service restart killed them halfway with no record; and the single run that completed end to end failed our own visual quality check.¹¹ The diagnosis was the useful part. The problem was never in the visually impressive final stage, it was upstream in camera alignment, and it took longer to find than it should have because the pipeline was not writing logs. That is a fair description of where WALLNUT stood with this technology in mid 2026: the capture is easy, the reconstruction is not. Nine runs of our own pipeline say nothing about anyone else's, and we are not going to extrapolate from them to the industry — but they are enough for us to ask a supplier who claims a fully automatic pipeline to show us the failed runs.

What it genuinely delivers

None of the above cancels the value, it defines it. What a client gets is a dated, revisitable record of the building as it stood on a given day, in open formats that stay with the project. Questions that used to require a site visit and a tape measure can be answered from a desk, weeks later, by someone who was never there. Design decisions on an existing building start from evidence rather than from a drawing that may be forty years old. And when there is a dispute about what was behind a wall before it was closed up, there is a file with a date on it.

Seeing Projects From Above

Drone photography feeds the same photogrammetric process, from above, and does the same job: a dated record of the site, produced quickly and without stopping work.

The practical limits are regulatory as much as technical. In the European Union, drone operations are governed by Commission Implementing Regulation (EU) 2019/947, which divides flights into open, specific and certified categories. Open category flights need no prior authorization but must stay within visual line of sight and below 120 meters, higher risk operations require an operational authorization based on a risk assessment, and Member States may declare geographical zones where flights are conditioned or prohibited entirely.¹² Add weather, and a flight is something to be planned rather than assumed.

The point that matters most to a client is not in any regulation: the accuracy of an aerial survey product comes from control, not from the aircraft. Unless the reconstruction is tied to points whose coordinates are independently known, it is a scaled picture rather than a survey.⁸ That is the question worth asking any supplier who presents an aerial model with dimensions on it.

Visualizing Before Building

One of the enduring difficulties in construction is that most people cannot read a set of drawings and see a room. Walking a model, on a screen or in a headset, changes that. Clients stop reviewing a design and start experiencing it, which tends to produce better feedback, earlier, while changes are still inexpensive.

The constraints here are measurable, and we have measured them. Testing our own viewer on a Quest 3 headset in April 2026, we recorded 200,000 splats rendering at 36 frames per second.¹³ The working budgets WALLNUT uses — around three million Gaussians for mobile devices and five million for desktop — come from our own testing, not from a published standard.¹³ A whole building at full capture density exceeds those numbers comfortably, which is why walkthroughs of complete sites need level of detail streaming rather than serving the whole file and hoping. That is unglamorous engineering, and it is the difference between a demonstration and something a client can open on an iPad on site.

Virtual reality also has a limit that is easy to forget while wearing a headset. It communicates space, sequence and proportion very well. It does not prove a dimension, and it does not reproduce real light or real materials. Decisions about how something fits go back to the model.

The Value of Integration

While each of these technologies offers real benefits, none creates much value alone. Drones, scanners, BIM software and headsets are now widely available across the construction industry, and owning them is no longer a differentiator. The advantage lies in how the information they produce moves through the project.

This is where conversations about construction technology usually stop short. What remains uncommon is a workflow in which information genuinely flows from design to engineering to construction without being re-entered, re-drawn or re-interpreted at every handover. Clients should not ask whether a company uses these technologies. They should ask how information is shared, who interprets it, and what decisions it is actually used for.

At many firms the tools sit inside separate departments. Architects work on one platform, engineers on another, contractors on a third, and information is transferred repeatedly across those boundaries. Every handover is an opportunity for delay, inconsistency and loss. WALLNUT is organized differently. Under a single design and build contract, architecture, engineering and construction operate as one team, so the same capture that documents an existing building also feeds the model and the comparison of model against reality — and will feed the cost view when that view has an interface. That continuity comes from the contract, and the software follows from it.

Looking Ahead

Predicting the future of construction technology is difficult, and the honest position is that this field is moving faster than its standards. That is starting to correct. In February 2026 the Khronos Group, which maintains the glTF 3D asset format, released a candidate extension, KHR_gaussian_splatting, to store Gaussian splats in glTF, with ratification expected during the second quarter of 2026.¹⁴ When capture formats standardize, captures stop being a supplier's proprietary asset and become part of the client's permanent record. For a client, that matters more than any improvement in image quality.

Other things remain genuinely unsolved. Measuring precisely inside a Gaussian splat is still an active research problem rather than a product feature,³ and automated progress assessment from imagery still requires human verification before it can support a payment. The persistent misconception about digital construction is that technology replaces expertise. The opposite has been our experience: the more capable the tools, the more the outcome depends on people who know which output to trust, which to check, and which to discard.

Technology is Only Part of the Story

What shapes the future of construction is the way technology, people and collaboration come together to produce a better building experience. Digital tools can reduce uncertainty, improve transparency and strengthen communication, but only when they are held by a team that knows how to turn information into decisions.

The firms that lead this industry will not necessarily be those with the newest technology. They will be those that combine capable tools with experienced people and a process that does not fragment. When that happens, clients gain something more valuable than innovation: clarity that everyone is working from the same information, that decisions are being made in advance rather than in reaction, and that their project is being managed as one effort from concept to completion.

References

1. International Organization for Standardization, ISO 19650-1:2018, Organization and digitization of information about buildings and civil engineering works, including building information modelling (BIM). Information management using building information modelling. Part 1: Concepts and principles, https://www.iso.org/standard/68078.html. ↩
2. International Organization for Standardization, ISO 16739-1:2024, Industry Foundation Classes (IFC) for data sharing in the construction and facility management industries. Part 1: Data schema, https://www.iso.org/standard/84123.html. ↩
3. WALLNUT internal research note, "Reality capture and 3D Gaussian Splatting: capture, processing and web delivery," June 12, 2026 (on file). Consolidates verified findings on capture density, frame selection, deliverable formats and the current state of measurement in radiance fields. ↩
4. Niantic Spatial, "Scaniverse: 360 Camera capture guide," technical documentation, https://nianticspatial.com/docs/scaniverse/360camera/. Specifies 8K30 recording, a fixed 1/500 shutter indoors, a minimum 30 cm standoff from surfaces, loop closure at the end of the scan, and a printed calibration board to establish 1:1 metric scale. ↩
5. WALLNUT internal research note, "Photogrammetric alignment software: edition capabilities and automation limits," June 12, 2026 (on file), verified against the vendor's published edition comparison and product manuals. ↩
6. WALLNUT internal research note, "Gaussian splat training: tooling, profiles and hardware limits," June 12, 2026 (on file), including published third party benchmark timings on comparable GPU hardware. ↩
7. WALLNUT internal research note, "Operator removal and mask handling in 360 capture pipelines," June 12, 2026 (on file). ↩
8. COLMAP, "Frequently Asked Questions: geo-registration," https://colmap.github.io/faq.html. A reconstruction is aligned to a real world coordinate frame by supplying known 3D locations for a subset of the images, from which the similarity transform, including scale, is estimated. ↩
9. Ivana Petrovska and Boris Jutzi, "3D Gaussian Splatting Methods for Real-World Scenarios," ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences X-G-2025 (ISPRS Geospatial Week 2025, Dubai), 641 to 648, https://doi.org/10.5194/isprs-annals-X-G-2025-641-2025. Table 1: RMSE against ground truth mesh, indoor scenario, MVS 1.43 mm, 3DGS-Basic 4.74 mm, 3DGS-MCMC 5.80 mm, Splatfacto 7.28 mm; vegetation scenario, MVS 3.23 mm, 3DGS-Basic 7.49 mm, Splatfacto 13.80 mm, 3DGS-MCMC 17.36 mm. ↩
10. Martyn Day, "Introducing Gaussian splats for AEC," AEC Magazine, December 2, 2025, https://aecmag.com/technology/introducing-gaussian-splats-for-aec/. ↩
11. WALLNUT internal pipeline audit, "Reality capture processing runs: failure analysis," June 12, 2026 (on file). Nine recorded runs, two terminating in error, one after 18.13 hours; further runs left in an unresolved state by service restarts; the single completed run failed internal visual quality assessment. ↩
12. Commission Implementing Regulation (EU) 2019/947 of 24 May 2019 on the rules and procedures for the operation of unmanned aircraft, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32019R0947. ↩
13. WALLNUT internal measurement, headset rendering test, April 8, 2026 (on file): 200,000 splats at 36 fps on Meta Quest 3. Device budget guidance for 2026 recorded in the same internal research note as ³. ↩
14. The Khronos Group, "Khronos Announces glTF Gaussian Splatting Extension," press release, February 3, 2026, https://www.khronos.org/news/press/gltf-gaussian-splatting-press-release. ↩