Trunk Tools Ditches General-Purpose LLMs To Tame Construction’s “Ugly” Data

Construction project management startup Trunk Tools is pushing back against the idea that a single, general-purpose large language model can handle the messy reality of enterprise data. Instead, the company says it has cut document review cycles from roughly two months to about 10 days by building a domain-specific AI stack tuned to construction workflows and documents.

The move reflects a broader tension in AI adoption: powerful foundation models excel at broad, conversational tasks, but often stumble when asked to reason over jargon-heavy, irregular, and proprietary data that underpins real-world industries.

A three-layer stack for “ugly” industry data

Trunk Tools describes most industry data environments as the opposite of clean SaaS dashboards. In construction, information is scattered across long-running projects, inconsistent formats, and legacy systems, with “ugly documents, proprietary schemas, [and] implicit workflows” that challenge off-the-shelf models.

To address this, the company has built a three-layer architecture perception, semantics, and agents designed specifically for construction project management and automation. Rather than relying on a single general-purpose LLM, Trunk Tools structures and enriches its data first, and only then trains AI models on top.

“We really set out to take the data from dispersed systems, pre-process it, structure it, go through our ontology into a knowledge graph, and then train AI models,” said Sarah Buchner, Trunk Tools’ founder and CEO and a former carpenter.

According to the company, this stack enables:

• Review cycles to shrink from months to days
• Prevention of costly errors in the field
• Autonomous agents that can reason over millions of pages of industry documentation

While Trunk Tools focuses on construction, it argues that this blueprint data pre-processing, an explicit ontology, and a knowledge graph feeding specialized models and agents can be replicated in other verticals wrestling with similar data chaos.

Why general-purpose LLMs stumble on niche workloads

The company’s approach speaks to a growing recognition that foundation models are not always sufficient for critical, domain-heavy workloads. General-purpose LLMs are trained on vast amounts of internet-scale data and optimized for breadth of capability rather than depth in any one area.

“General-purpose LLMs are trained to be okay at everything, so they’re weak at anything niche,” said Kriti Faujdar, a senior product manager working in AI infrastructure, agentic AI, security, and LLM platforms. That weakness shows up around rare terminology, highly specialized reasoning, and the unspoken assumptions that experienced practitioners take for granted.

Developer Sébastien De Bollivier points to reliability issues when models face dense technical language and rigid formats. He describes the biggest bottleneck as performance on data that is “jargon-dense, abbreviation-heavy, and format-specific.”

As an example, he notes that “a GPT-4-class model can understand a French legal contract, but will fumble the specific article references practitioners need to cite.” The model can parse the language, but may miss the precise, citation-level accuracy that legal professionals require.

Compounding the problem, much of the most valuable enterprise data internal documents, proprietary formats, unique workflows never appears in the public datasets used to pretrain foundation models in the first place. Without targeted adaptation and structure, even state-of-the-art general models can struggle to deliver the reliability and context that industry teams need.

Trunk Tools’ experience suggests that for sectors like construction, the path forward may lie less in waiting for bigger general-purpose models and more in reshaping messy, fragmented data into industry-aware knowledge structures then pairing that with specialized AI agents.

Source: VentureBeat