Construction AI Reduces Document Review from 60 Days to 10 Days

Trunk Tools, a construction project management company, developed a proprietary three-tier architecture specialized for the construction industry without relying on general-purpose AI models. According to the company, this effort successfully reduced document review periods from several months to just days. The company also reports that it has contributed to preventing costly mistakes on construction sites, and AI agents have become capable of autonomously interpreting materials spanning millions of pages.

In the construction industry, large volumes of paper documents and proprietary format data—such as blueprints, specifications, and contracts—are scattered across various locations, and most construction sites lack organized databases. While general-purpose large language models (LLMs) like ChatGPT can handle a wide range of applications, they struggle with industry-specific terminology, implicit conventions, and unique formats. Sarah Buchner, CEO of Trunk Tools and a former carpenter, identified this challenge and undertook the development of an industry-rooted system.

The system built by the company consists of three layers: "perception," "semantics," and "agents." CEO Buchner explains: "We collected data from fragmented systems, preprocessed it, structured it, converted it into a knowledge graph through our proprietary ontology, and then trained the AI model." This series of processes enables highly accurate information processing suited to practical construction site operations.

Regarding why general-purpose LLMs become weak in niche fields, Kriti Faujdar, a senior product manager at an AI infrastructure company, points this out succinctly: "General-purpose LLMs are trained to handle everything adequately, which makes them weaker in specialized domains." He further notes that a company's most critical data—that which lies dormant in internal systems and proprietary formats—was never included in the model's pretraining in the first place. While RAG (Retrieval-Augmented Generation: a technique that generates responses while referencing external data) is somewhat effective, he states that it "merely provides better facts to a model that is fundamentally weak at domain-specific reasoning."

Faujdar emphasizes that in implementing industry-specialized AI, priorities should include first conducting pretraining with industry-specific data, followed by fine-tuning with practical operational data, and then building proprietary evaluation metrics. He also states: "Thousands of data points prepared by actual domain experts are more valuable than millions of noisy scraped data points." Additionally, web engineer de Bollivier argues that "fine-tuning for reliable output formats required in workflows is more important than fine-tuning to make the model smarter in the domain," and points out the effectiveness of hybrid configurations combining general-purpose models with specialized smaller models.

The Trunk Tools case demonstrates that in achieving industry-specialized AI, "how data is organized and trained" creates a more fundamental difference than "which model is used." While general-purpose LLMs appear to serve all industries, use cases requiring field-level accuracy and reliability necessitate such specially designed architectures. Beyond construction, the company's approach is positioned as a reference model for designing "industry-specialized stacks" that structure dispersed data for AI utilization across various sectors.