In the construction industry, document management is a logistical nightmare: complex projects, millions of pages of drawings, specs, and contracts, with review cycles measured in months. Trunk Tools, a construction project management company, decided to break this pattern by abandoning general-purpose AI models and building a proprietary three-tier architecture based on perception, semantics, and agents. The result is a reduction in review cycles from 60 to just 10 days, with significant cost and time savings.
Why general-purpose models fail in specialized domains
Traditional large language models are trained to be good at everything, but precisely because of that, they fail in highly specialized fields. Kriti Faujdar, a senior product manager specializing in AI infrastructure, explains that tacit knowledge of industry practitioners, rare terms, and technical abbreviations are insurmountable obstacles for a GPT-4-class model. "A generic model can understand a French legal contract, but stumbles on the specific article references a lawyer needs to cite," adds developer Sébastien De Bollivier. The most valuable enterprise data never makes it into public pretraining. Large language models provide a foundation, but for vertical precision, custom architectures are necessary.
Sponsored Protocol
In this context, the AI chip shortage pushes many companies to seek more efficient solutions, avoiding the computational overhead of generic models. Trunk Tools took the opposite path: instead of adapting an existing model, it built a modular system that leverages deep learning only where needed.
Trunk Tools' three-layer architecture: perception, semantics, agents
CTO Amrish Kapoor explains that probabilistic models are insufficient for construction: a two-millimeter symbol in a drawing can have vastly different meanings depending on its placement. The first layer, perception, teaches AI to read the symbolic language of blueprints, recognizing doors, beams, and other elements not always explicitly labeled. The second layer, semantics, builds a knowledge graph that connects data: a door is not just an arc on a wall but is linked to the drawing detailing it, the spec governing it, and the trade installing it. On top of these two layers operate AI agents that perform specific tasks such as reviewing requests for information or analyzing bids. Each agent achieves 95% accuracy before release, thanks to continuous evaluation pipelines and an LLM-as-a-judge system that measures both objective accuracy and subjective quality of generated responses.
Sponsored Protocol
Measurable results: from 60 to 10 days with quantifiable savings
The submittal review agent, for instance, compressed the cycle from 50-60 days to 10 days, with direct impact on construction schedules. In one concrete case, the agent detected that a structural beam had been moved up 8.5 inches without the architect documenting it: an error that, if not caught during design, would have cost over $10,000 in rework. Other examples include flagging $60,000 in inflated pricing from a landscaping subcontractor and identifying a fireplace that needed sealing before drywall installation, saving about $100,000. Customers report average savings of 8 minutes for single document retrieval, 20 for standard referencing, and up to 75 minutes for complex tasks.
Sponsored Protocol
Lessons for other industries
According to Sarah Buchner, CEO of Trunk Tools and a former carpenter, this approach can be replicated in any vertical with unstructured data. The key is to understand the specific data challenges of the domain and build technical infrastructure that transforms information chaos into structured data for language models. "Build your technical advantage where generic models are not investing and not performing well," advises Buchner. European tech policies could draw inspiration from this specialization, favoring ecosystems that incentivize vertical solutions rather than uncritical adoption of generalist models. The path to industrial AI is paved with domain-specific data, not with omnipotent models.
