Advancing AI Capability for the Built Environment

Artificial intelligence has a spatial problem. The models that power modern AI were trained on text, code and images sourced overwhelmingly from the internet. The investment in their capabilities has been directed towards problem areas where that data is rich: software development, legal analysis, financial services, content generation. These are domains where the core intellectual work happens in language and logic, and where the training data to build capable systems has been abundant and accessible for years.

The built environment sits outside this.

Structural engineering, architecture, civil engineering, building services: the knowledge that drives these disciplines is spatial, physical and three dimensional. It lives inside proprietary analysis models, BIM geometry, load calculations and design coordination processes that have never been published to the internet and never will be. The professionals who hold this knowledge think in forces, materials and spatial relationships, not in paragraphs.

AI capability has advanced rapidly, but it has advanced along axes that do not intersect with how buildings are designed. A large language model can discuss structural theory with fluency because it has read the textbooks. It cannot operate within a structural engineer's actual working environment because that environment is composed of software, data formats and spatial reasoning that language models were never built to process.

Private data: The most valuable knowledge in a structural engineering firm is the most private. How a firm approaches a complex transfer structure, how they have refined a connection detail over 15 years, how a senior director recalls a comparable foundation solution from a project completed in 2008: this is proprietary, commercially sensitive and in many cases safety critical. It has no presence in any training corpus. Firms will not and should not send it to a third party model provider for training.

Incompatible formats: Even if the data were available, it is not text. An ETABS analysis model encodes the structural behaviour of a building as a three dimensional system of forces, stiffness and geometry. A Revit model encodes spatial relationships, material properties and construction sequences. These are structured engineering artefacts that require specialised software to interpret. A language model trained on their raw file contents would learn nothing meaningful because it has no framework for understanding what the data represents in physical space.

No margin for error: Language models produce plausible outputs, not verified ones. In software development, a plausible but incorrect suggestion is caught by a compiler, a test suite or a code review. In structural engineering, a plausible but incorrect design decision affects the safety of a building that people will occupy for decades. The feedback mechanisms that make AI useful in software do not exist in building design, and the consequences of error are in a different category entirely.

We asked a large language model directly: why do you struggle to understand the built environment and why can you not streamline workflows in proprietary engineering file types?

The answer was precise: The model identified that it has no true spatial understanding, that its knowledge of the built environment is derived entirely from text rather than engagement with three dimensional models. It stated that it operates statelessly and cannot maintain the persistent, structured model state that engineering workflows depend on. It acknowledged that proprietary file formats are closed, complex and dependent on specific software ecosystems that it cannot access. And it noted that it generates probabilistic responses where engineering demands deterministic precision.

The model's own conclusion: "The primary limitation of LLMs in this context is not a lack of conceptual understanding, but a lack of integration with the systems that define engineering practice."

It then identified its most useful role: translating design intent into scripts and automation logic, explaining software APIs, assisting with data transformation and advising on system integration. In other words, the model described itself as a natural language interface waiting for an engineering execution layer to connect to.

That execution layer is what we are building.

Bite builds proprietary tools that natively understand engineering software. Bite can read the data inside analysis and modelling packages, interpret the relationships between structural elements, detect changes to a design and understand their consequences across the full tool stack. This is domain specific engineering intelligence, built through direct integration with the software that engineers use every day.

The role of the large language model in our architecture is precisely defined. It captures engineering intent expressed in natural language and translates it into structured commands. Bite execute those commands against the engineering software. The model handles language. Bite handles engineering. The engineer makes the decisions.

We are building Bite to interoperate through the Model Context Protocol, an open standard for connecting AI models to external tools and data sources. This creates two independent layers that evolve separately. As language models improve at understanding human intent, the conversational interface improves. As Bite deepens integration with engineering software, the execution capability improves. Neither layer depends on the other for its core function.

Our customers' engineering data stays within their own environment. It is never extracted for model training. It is never shared across clients. It is never used to improve a foundation model. When our tools operate on engineering data, they do so in place.

Structural engineering firms hold data that represents decades of practice, hundreds of millions of pounds of delivered work and the institutional memory of the organisation. The data is their competitive advantage and in many cases is subject to contractual confidentiality. Any technology that operates within building design must treat this data with the same seriousness that the firm itself does.

The built environment is one of the largest sectors of the global economy and one of the least affected by the current wave of AI capability. The reason is structural: the data is private, the formats are specialised and the tolerance for error is fundamentally different from the domains where AI has gained traction.

Larger language models will not close this gap. More training data will not close it either. The gap is not about what AI knows. It is about what AI can reach.

Today, AI capability sits on one side: language, reasoning, pattern recognition. Engineering sits on the other: spatial, physical, embedded inside proprietary software that AI has no way to interact with. An engineer cannot take the reasoning power of a language model and apply it to the data inside ETABS or Revit because the model has no mechanism to read, interpret or act on that data.

Bite is that mechanism. We create a machine readable interface to engineering software, translating the spatial, three dimensional world of building design into structured operations that AI models can participate in. LLMs captures the engineer's intent in natural language. Bite translates that intent into precise commands that execute against the engineering software: reading analysis results, modifying structural elements, detecting downstream impacts, propagating changes across the tool stack.

Intelligence flows through Bite. The LLM provides the conversational surface where the engineer expresses what they want. Bite provides the engineering surface where that intent becomes action. Together, they create a path for AI capability to operate within building design for the first time.

We are starting with structural engineering, where we have deep domain expertise and validated understanding of the problem and extending across the disciplines involved in designing the built environment.

Bite Engineering is an AI Infrastructure and Services company.