Audio article narrated by OpenAI
The missing piece in construction AI isn’t a better model. It’s an interface: a structured format that lets AI reason about buildings as buildings actually work, in three dimensions, under load, within code. Without it, every spatial AI system in construction is solving the wrong problem. Endra started by solving the right one. I sat down with Niklas to understand what it took to build a spatial AI in construction. What follows is a conversation about first principles, the economics of speed, and what happens to a profession when the tool finally matches the complexity of the work.
About Niklas Lindgren
Niklas is the CEO and co-founder of Endra, a deep-tech platform for MEP (mechanical, electrical, and plumbing) design. A practitioner first, Niklas has been a Revit user since 2013 and spent most of his career in the MEP engineering world before co-founding Endra with Anton and a technical team whose roots are in Goldman Sachs’ low-latency trading infrastructure.
“A building system isn’t just geometry. Behind every cable, duct and component are voltage drops, load calculations, and code constraints. At Endra, we generate the 3D models, but they’re always driven by the physics of the system.” Niklas Lindgren
Endra’s seed investor did their homework before committing: 40 to 50 customer reference calls. The firm Notion Capital specialises in vertical B2B software and has no particular expertise in AEC. Yet when the calls came back, Niklas says the partner told him they had “never had anything like that type of response from users before.” The conviction was enough to back a company that would burn through tens of millions in R&D before reaching scale.
What made those conversations so unusual? Partly the team. Half of Endra’s founding group came from Goldman Sachs’ low-latency trading infrastructure: engineers who built systems where milliseconds matter and correctness is non-negotiable. The other half, including Niklas and Anton, came from years of MEP practice. That combination is rare; the investor bet it was rare enough to matter.
Partly, though, it was the problem itself. MEP, the mechanical, electrical, and plumbing systems that make a building habitable, sits in an odd position in the AEC world. It represents a significant share of the project cost and is often where engineering firms make the most money. Yet it has received almost none of the software investment that has gone into architecture and structural engineering. Niklas is direct about why incumbents have stayed away: “It’s a very large risk and it’s very expensive.” Endra planned from the outset to invest over $40 million in R&D across three years. “That has never been done in MEP software before.” To do this, Endra hires differently from most AEC software companies: PhDs from Princeton, engineers who worked on simulation software for autonomous vehicles, and people with deep-tech backgrounds. “We’re pushing beyond what’s historically been possible – both in industry and in academia. To do that, we need the smartest people on the planet on our team.” Niklas discusses the research at Endra. “It’s a people’s business. Building a software organisation is most and foremost a people’s business.”
And then there is the CTO, David Rydberg. A KTH Royal Institute of Technology alumnus who went on to build equity trading infrastructure at Goldman Sachs, Rydberg took a detour before founding Endra: he built and open-sourced a local large language model, one designed to run on a personal computer, developed at his Stockholm apartment. This was 2023, one year before DeepSeek brought the concept to mainstream attention—no press coverage, no public repository; just a working model that circulated in technical circles. The depth of in-house AI capability was built before the wave arrived.
Endra started with the architecture.
The missing layer between language and space
The dominant narrative around AI in construction goes something like this: ingest drawings, apply a model, generate outputs. The problem is that between the natural language layer (what LLMs do well) and the physical world (what buildings are), there’s no native interface. Text can describe a floor plan. It can’t reason about whether a sprinkler head creates a double-coverage conflict, or whether a smoke detector should sit adjacent to an air inlet.
This is the gap Endra built into. The company created a proprietary 2D/3D geometry engine from scratch and a custom file format, EDRF, designed specifically for LLMs to reason about spatial data. Niklas is clear that this format is part of Endra’s core defensibility: “We see it as part of our moat and part of our defensibility as a firm.” It functions as a kernel: a structured intermediate representation that translates between the geometry of buildings and the language-based reasoning of AI.
The World Labs framing of “3D as code” captures why this matters: just as text became the universal interface between humans and software, 3D structured formats are becoming the interface between humans and AI agents operating in physical space. The key property isn’t visual fidelity; it’s composability. A structured spatial format is inspectable, editable, and consumable by other systems. Pixel outputs are not.
Niklas has tried Marble, World Labs’ generative world model, and is precise about where the two approaches diverge. “World Labs will have a huge impact on creating interior concepts for hotel rooms,” he says. “MEP is something that’s not super artistic and has a lot of constraints. You can’t just design for it to look nice. You have to design with voltage drops in mind, with load calculations in mind, with building code.” His longer-term ambition builds on that distinction: Endra aims to let engineers generate full MEP concepts within physics and code constraints, inhabiting the same three-dimensional world as generative spatial AI, but with hard engineering rules as the boundary conditions.
Endra’s kernel is that structured format: expressive enough for LLMs to reason over, deterministic enough to carry physics, and editable enough for human engineers to intervene.
Three layers that can’t be collapsed
What Endra built isn’t a single AI model doing everything. It’s three distinct layers working together, and the discipline of keeping them separate is what makes the system trustworthy.
The first layer is probabilistic. “LLMs play a very important role on our platform,” Niklas explains. “Among many things, the most obvious is that LLMs help us understand the context of a building — what type of room it is, how it’s used, and what the expected occupancy is. From there, we can match that information to building code and the designer’s intent.” This is where language intelligence applies: where the system understands that a room labelled “ICU” implies different requirements than one labelled “conference room.”
The second layer is deterministic. Engineers encode constraints: placement heights, mounting preferences, clearance rules specific to their firm’s standards. These are not things that should hallucinate. A smoke detector placement that violates a mounting rule isn’t a suggestion; it’s a defect. This layer is user-controlled, explicit, and fixed.
The third layer is physics-based. Cable sizing, lux level calculations, load calculations, breaker sizing: hard constraints governed by engineering principles, not approximations. These feed into the model but are never overridden by probabilistic outputs. Everything connects into what Niklas calls “one closed data model, which is very tight and works very well.”
Francesco Iorio, writing in Forbes Tech Council, identified exactly this requirement: spatial AI for construction needs three things that can’t be shortcut: decades of construction intelligence, real-world project datasets, and a deterministic physics engine. “In construction, the non-deterministic nature of AI can mean critical design errors that cascade across an entire project.” A hallucination in a blog post is a typo. A hallucination on a construction site is a $2 million change order.
The three-layer architecture is what prevents the latter. Endra’s LLM never touches physics. The physics engine never speculates. And the engineer retains full visibility and control over both.
When 400 hours become eight
The practical output of this architecture is hard to accept without seeing it. Endra ingests linked models from every trade: HVAC, plumbing, sprinklers and structural. It auto-annotates the entire model and then reasons spatially across the entire model.
Smoke detectors are placed between beams, not adjacent to air inlets. Sprinkler heads positioned to avoid double-coverage conflicts. The system knows the geometry, the constraints, and the relative positions of the physical objects. “Once you give Endra the context,” Niklas says, “it will, to a very high degree, like 90% solve a lot of the issues that you have. And then it will flag everything that Endra doesn’t solve.”
The time compressions are the consequence: 30 hours of coordination reduced to 1.5 hours; 400 hours of MEP design completed in a single day; 300 hours of electrical design condensed into 30 minutes. These come from closed deals and measured delivery results, not projections.
For MEP firms, this speed was previously unattainable because it required holding the entire model in a single system’s awareness simultaneously. That’s what the kernel provides. Not faster drafting. A different class of reasoning.
When speed breaks the pricing model
Here is where the technical story connects to the commercial one, and where the implications extend beyond Endra’s own business.
If an MEP firm historically billed 400 hours to produce a full design package, and Endra compresses that to eight hours, the billable-hour model produces a perverse outcome: the firm earns less for better work. That’s the fear that has made automation a threat, not an opportunity, for professional services firms in AEC.
” Commercially, Endra exists for one reason: to help our customers make more money dramatically. They increase margins on their projects, offer more attractive and faster products and scale project volume with us. If we don’t achieve that, we’ve failed – and we have no right to exist.”
Niklas describes two recent conversations that reframe the question. The first: a top-tier MEP firm in New York told him they believe they can raise fees to property owners by delivering designs faster. “If they can hand out their designs faster, then they probably will be able to charge more to their customers while spending less time.” The second came by accident, during a separate conversation with one of the largest real estate owners in New York: he confirmed he “would be open to pay more for something that came to him faster.”
The economic logic is not complicated, but it’s easy to overlook. A developer who has acquired a Manhattan site is losing significant money every month while waiting on design approvals before leases can start. If MEP design timelines compress from 12 months to two months, and lease revenue runs at $1 million per month, the developer recovers $10 million. In that context, paying more for MEP design isn’t a concession; it’s an obvious trade.
The same logic applies differently in data centres, a vertical Endra is actively targeting. “Data centres have a different reason to push design timelines,” Niklas says. “It’s competition. If they can get their data centres online, they can start selling compute faster, hence own more market share.” The urgency is different from real estate, but the outcome is the same: speed is not a nice-to-have; it’s a competitive weapon.
The value isn’t the hours. The value is the outcome: faster occupancy, earlier revenue, reduced financing costs.
Not every market will move at the same pace. In many regions and many client relationships, the default remains cost reduction: clients expect automation to lower fees, not raise them, and they are not wrong to try. The more durable version of the argument is about who captures the productivity gain, not whether one exists. A firm using Endra on a fixed-fee project doesn’t have to pass every saved hour back to the client. It can price competitively, win work it previously couldn’t, and maintain margins. Others will charge more where the market allows it. The goal isn’t a single pricing model; it’s a delivery platform that lets a firm compete on whatever the client values most: speed, cost certainty, quality, or innovation. The margin follows either way.
This connects to a shift visible across professional services. On a McKinsey podcast, Naveen Chaddha, managing director at Mayfield Fund, made the same argument: “Don’t charge for the number of hours. Don’t charge per seat. Charge for the work you do and the outcome you create.” Legal firms face the same pressure; the billing model built on input time is increasingly misaligned with where AI actually creates value.
Niklas notes that in the US market, fixed-fee pricing is already more common than the billable-hour model (which remains more prevalent in Europe). Firms Endra works with in the UK are already moving toward fixed pricing. The direction of travel is clear: when AI collapses effort while maintaining the value of the output, input-based pricing becomes structurally incoherent.
Codifying what engineers know
One of the less-discussed implications of this architecture is what it makes possible beyond speed. Endra’s next major product, to be released in Q3 this year, is a system that lets firms encode their institutional know-how into reusable digital workflows. Niklas is not providing all the details but is explaining the concepts.
Every MEP firm carries knowledge that lives in people’s heads. Senior engineers who know that a particular building type always needs a certain configuration, or that a specific client expects devices positioned in a particular way. When those engineers retire or move on, that knowledge goes with them. Industry data projects that 41% of the construction workforce will retire by 2031: the knowledge drain isn’t hypothetical; it’s already scheduled.
Niklas draws the parallel to how software teams now codify intelligence. “Taking from another industry, in this case, the software industry, where skills and reusable workflows came from: the concept of saying we have so much knowledge that’s in the hands of someone, or in Excel spreadsheets, and we can now pack it up in very readable, plain text that can be reused.” The next product applies the same logic to MEP expertise. The difference is that what gets codified isn’t a programming pattern; it’s decades of project-type knowledge that currently exists nowhere except in the heads of experienced engineers.
This will also have a meaningful competitive consideration. A firm that has encoded years of project-type expertise into Endra has created something a competitor cannot replicate by switching software. The knowledge becomes a durable asset, not just a workflow.
A moat built from the ground up
Endra’s design choices signal something consistent throughout: the system is built for engineers who need to trust it fully, not users who can tolerate occasional errors.
The human-in-the-loop principle isn’t incidental. “We’re not building it for autopilot and just click a button and see what happens,” Niklas says. Engineers can edit everything on a Figma-style 2D/3D canvas; the LLM proposes, and the engineer controls the outcome. In Niklas’s estimate, there will always be 5–10 per cent of every project that is genuinely custom: a client calling to say they want 12 receptacles in a specific room, or a particular fire alarm system with no data equivalent. “That’s where the engineer will add value. And I think there will be lots of those discussions.” The engineers Endra works with aren’t scared of this shift; they’re largely relieved to hand off the drudge work, placing thousands of outlets and running cable schedules, and focus on the decisions that actually require judgment.
This matters particularly in MEP, where errors cascade. A wrong cable size doesn’t fail in isolation; it creates a chain of downstream problems across the whole model. The architecture’s determinism is a trust mechanism as much as a technical one.
The interface that was missing
The deeper significance of what Endra is building isn’t the speed, though it’s remarkable. It’s the demonstration that spatial AI in construction requires a different architecture than general-purpose AI: one built from physics up, not from language down.
The spatial kernel, the structured intermediate representation that makes LLM reasoning about buildings possible, is the piece that’s been missing. Once it exists, the downstream effects compound: faster delivery enables new pricing models, which attract better clients, which generate better project data, which makes the models more accurate, which makes the kernel more valuable.
And the implications stretch further still. “Connecting the model to the real world – whether for manufacturer configuration, prefab, or robotics – starts with a deeply granular representation of the built environment,” Niklas says. “And we can provide that.” The kernel isn’t just a design interface. It may be the foundational data layer that one day connects the digital model to the physical installation entirely.
Endra’s technical foundation is also its commercial thesis: speed is the product, outcomes are the price, and the firms that move first will carry institutional knowledge their competitors can’t buy.