Higharc just demoed a one-shot pipeline that turns a hand-drawn floor plan into a fully buildable home: BIM model, structural framing, construction documents, cost estimates, and 3D renders, all from a single AI prompt. The breakthrough is treating a building like language. A transformer model predicts the next wall, door, or window the way ChatGPT predicts the next word. It only works for homebuilding today. The playbook is portable.
What used to take months just happened in minutes. A team of architects, engineers, estimators, and renderers passing files between half a dozen tools, redrawing the same lines they have drawn a thousand times before. Higharc compressed that into one prompt.
Hand-drawn napkin sketch to fully buildable home. Construction documents, structural framing, cost estimates, renders. We have been speculating about speech-to-design on the podcast for a couple of years, asking guest after guest when it would actually arrive. The answer, apparently, is now.
The problem that nobody has cracked until now
Why every project still starts from zero
To understand why this matters, you have to understand what is broken about how buildings get designed today. Vaidas, an AI engineer on Higharc’s special projects team, started his career as a structural engineer at Arup. He frames it like this. Imagine you have designed a thousand data centres. You have all the knowledge in your head about how to do it. But when you sit down to design your thousand and first, your tooling does not reflect any of that. You still draw every line in the BIM model. Or you copy and paste something from a different job and start cleaning it up.
Every project starts from zero. Decades of expertise live in the architect’s head, not in the tools. That gap is where the time and cost bleed out. It is also why generic generative design tools have produced famously bad outputs, the floor plans with three sinks in the master bathroom, the staircases leading nowhere. The AI was not the problem. The data model was.
Every project starts from zero. Decades of expertise live in the architects’ heads, not in the tools they use.Vaidas, AI Engineer, Higharc Special Projects
The demo: sketch, prompt, buildable home
An end-to-end pipeline in one go
The team walked us through the pipeline founder Marc Minor and co-founder Michael Bergin have been building. A user uploads a hand-drawn floor plan. Higharc’s computer vision model reads the lines, identifies the rooms, and parses everything into the platform’s native BIM-like data model. Within seconds you have walls, doors, windows, and structural framing. Because that data model is connected to every other layer of the system, 100% construction drawings appear on the right-hand side of the screen as it builds out. Cost estimates follow automatically because the structural framing is real, not decorative.
Then comes the part that has been a research dream for years. The user types into a chat panel: move the garage wall out four feet. The wall pops out four feet. Same chat, next prompt: render this house on a lakeside. Diffusion model fires, photorealistic image lands. Same chat: now do it as a watercolour. Different render. Sketch to BIM to construction drawings to renders to live edits, all in one continuous conversation.
- Reads the sketch. Computer vision parses the floor plan, identifies rooms, tags spaces, and matches them to Higharc’s defined data types.
- Builds the model. Walls, doors, windows, structural framing and roof timbers are generated as real, buildable entities, not just geometry.
- Pulls the deliverables. Dimensioned construction drawings, cost estimates and 3D renders fall out of the same model, no separate tools required.
- Takes voice or text edits. Natural-language prompts adjust walls, extend porches, or restyle the render, with the model re-coordinating downstream.
The breakthrough: a building is a language
Why a transformer beats every previous attempt
The deceptively simple idea came from Manuel, the architect-turned-machine-learning engineer who cracked the core problem. He looked at the way language is structured and realised buildings work the same way. Text has a body, paragraphs, sentences, words, characters. A home has levels, areas, rooms, then entities like walls and doors inside those rooms. A hierarchical system either way. So Manuel built a transformer, the same family of model that powers ChatGPT, but instead of predicting the next word in a sentence, it predicts the next entity in a building. He found a way to tokenise a home.
This is the moat. Plenty of ML researchers have poked at this problem. Few of them are also architects. Manuel is one of the small number of people in the world with both, which is why Higharc has shipped what others have only sketched out in papers. AEC Magazine’s Martyn Day, who covered the September 2025 beta in his piece, noted the model even auto-generated structural timber roof framing that was never drawn. He called it gobsmacking. That is not a word AEC press uses lightly.
Generic AI floor-plan tools work on pixels and shapes. Higharc’s model works on a data structure that knows what a kitchen needs, what a bedroom needs, and how walls connect to roofs. The AI is not drawing pictures of houses. It is autocompleting a building.
What this unlocks for home builders
Anyone who can read can now design a home
Higharc is purpose-built for builders putting up 20 or more homes a year. The customer list, drawn from the company’s site, includes Tri Pointe, Buffington, Epcon, Eastwood, Eddy, and Crystal Creek Homes. Replace the patchwork of CAD, BIM, estimating tools, sales configurators, and rendering suites with one platform where every output derives from the same digital model. The sales team’s brochure, the field crew’s drawings, and the procurement team’s takeoff all reference the same source of truth.
Try to upload a bridge or a skyscraper and it will not work. Higharc has been clear about that. The depth comes from the constraint. Every US timber-frame home is a variation on a theme, and the team has drilled into the granularity of that theme until the AI knows what a kitchen wants before you tell it. The published research, though, maps to other typologies. Someone will run this playbook for data centres. Someone else for high-rises. Higharc got there first for homes, which is exactly where the cost-of-tooling pain meets the volume to justify the work.
This whole system allows people to design as naturally as they talk. We’re removing the expertise layer that complex software entails. Now everyone who can read can design their own homes.Manuel, Machine Learning Engineer, Higharc
Get the briefing construction execs actually read
Sharp takes on AI, design, funding, and the tools reshaping the industry. One email, no fluff.
Join 2500+ ReadersThe bigger picture: design tools are about to change
Expert systems vs general-purpose BIM
What Higharc is doing fits a broader shift the AEC software world is gearing up for. Andreessen Horowitz published a thesis earlier this year arguing Revit, the industry’s default BIM tool since 1997, is ripe for replacement. AI vision and language models can finally parse the unstructured metadata inside BIM files, which is what unlocks real design intelligence rather than just faster drawing. We covered the full a16z breakdown here.
Higharc sits in a slightly different lane. It is not trying to replace Revit. It is building an expert system for one building type and going incredibly deep on it. The open question for the rest of AEC is whether expert systems will fragment the market, with bespoke platforms for hotels, labs, data centres and warehouses, or whether one general-purpose AI-native BIM tool wins. The jury is out. What is not in doubt is that we are now talking to design tools, and the tools are talking back. Our AI agents guide covers the wider shift.
How Higharc stacks up against legacy design workflows
The same job, two very different processes
| Step | Traditional CAD/BIM workflow | Higharc AI pipeline |
|---|---|---|
| Floor plan input | Drafter recreates sketch in CAD, manually | Hand sketch uploaded, parsed by computer vision |
| 3D model | Built in BIM tool, walls and openings drawn from scratch | Generated from sketch, walls and openings inferred |
| Structural framing | Engineer designs framing, separate model or layer | Auto-generated, including roof timbers |
| Construction drawings | Drafted manually, redone with every revision | Pulled live from the model, always in sync |
| Cost estimates | Manual takeoff in spreadsheet or estimating tool | Auto-generated from the framing data |
| Design edits | Reopen file, redraw, push changes downstream | Natural-language prompt, model and drawings update |
| Renders | Export to rendering tool, hours per image | Diffusion model in same chat, seconds per image |
| Time to complete | Weeks to months | Minutes |
Frequently asked questions
Higharc is a live production platform used by builders including Tri Pointe Homes, Epcon Communities, Buffington Homes and Eastwood Homes. The full sketch-to-model-to-drawings pipeline shown in the latest demo is a research preview the team has been pushing into beta through 2025, building on the sketch-to-BIM capability they released earlier in the year. (Source)
US timber-frame production homes, the kind built by volume housebuilders putting up 50 or more units per year. Try to upload a skyscraper or a bridge and the system will not work. The depth of automation comes from the depth of the constraint. (Source)
Because the breakthrough requires both deep architectural knowledge and deep machine learning expertise, and almost nobody combines the two. Higharc’s lead ML engineer Manuel is a trained architect, and the model treats a building as a hierarchical language structure that can be tokenised. Earlier attempts at AI-generated floor plans worked on pixels and produced nonsense outputs like three sinks in one bathroom. (Source)
CEO Marc Minor came from 3D printing, having led Carbon3D and Desktop Metal. Co-founder and VP of Product Michael Bergin previously led the AEC research team at Autodesk and worked on Dreamcatcher, Autodesk’s generative design system. CTO Peter Boyer is a former Autodesk engineer and a founding member of Dynamo. The company has raised more than $75 million across Series A and B, with backers including Home Depot Ventures, Schneider Electric’s SE Ventures, Fifth Wall, Ferguson, Starwood Capital and Suffolk Technologies. (Source)
No. It means the parts of the workflow that have always been redrawing the same lines, generating the same takeoffs, and producing the same construction documents are moving into the tool itself. The expertise that lives in an architect’s head, what makes a kitchen work, how a house should sit on a lot, why one roof pitch beats another, becomes the layer the AI augments rather than replaces. For production homes, the tooling now reflects decades of cumulative knowledge automatically. (Source)
The research is portable, even if the current product is not. Higharc’s published methods map to other architectural typologies. Expect to see expert AI BIM systems emerge for data centres, hotels, labs and warehouses over the next few years. Each one needs the same depth of domain modelling Higharc has put into homes, which is significant work. (Source)
Revit and most general-purpose BIM systems require drafters to build everything manually, layer by layer. Higharc replaces static plans with generative digital models where every output, including drawings, estimates and renders, derives from the same source of truth. It is narrower in scope but far deeper in automation for the building type it covers. Andreessen Horowitz recently argued legacy BIM is overdue for AI-native replacement, which we broke down here. (Source)
Related Articles
The VC Firm Behind Facebook Just Told Construction: Your Software Is Next
AI Agents: The Definitive Guide
