Trunk Tools, a firm focused on construction project management, recently announced a new approach to document review. This breakthrough dramatically reduces processing time, challenging the current reliance on general-purpose AI. The company developed a custom stack integrating perception, semantics, and agents. This architecture handles the complex, messy data found in fields like construction and law. Trunk Tools claims this custom design overcomes the weaknesses of broad AI models. These general models often struggle with highly specific, jargon-filled enterprise information. This new method could serve as a blueprint for other sectors facing massive amounts of proprietary documentation.
Key Details
Foundation large language models (LLMs) offer powerful tools, but they lack niche understanding. General-purpose LLMs train well across many tasks, yet they perform poorly when facing specific domain problems. Sébastien De Bollivier, a software developer, noted that reliability remains the biggest bottleneck. This occurs when models confront data that is heavily abbreviated or format-specific. A GPT-4 class model might understand a French contract, but it may miss citations essential for a professional’s work. Furthermore, Kriti Faujdar, the senior product manager, pointed out a major issue. The most valuable enterprise data never enters the initial pre-training process. This critical information remains trapped inside internal systems and various proprietary formats. This severely limits the general model’s ability to reason correctly in the field. Retrieval Augmented Generation (RAG) helps provide facts, but it cannot fix a model lacking true domain-specific reasoning. Therefore, companies should fine-tune models using real task examples from professionals.
How Custom AI Solves Complex Data
Trunk Tools built a purpose-made, three-layer stack to beat standard LLM limitations. Construction workflows inherently contain implicit assumptions and connections. This information often scatters across many different sources. Sarah Buchner, the company’s founder and CEO, stated that project data is simply too vast for humans alone. The specialized system achieves high accuracy and relevance in its automation. This customized approach successfully shrunk the document review cycle from months to just ten days. The system also helps prevent expensive mistakes later in the construction field. This specialized stack gives autonomous agents the ability to reason over millions of pages of project documentation.
Trunk Tools’ architecture addresses the limits of probabilistic models. These models often fail at high-precision symbolic interpretation in complex documents. Amrish Kapoor, the company’s CTO, explained that standard transformers might only identify a tree. They cannot interpret a 2-millimeter symbol that means something else depending on its placement. The three layers of Trunk’s system break the workflow into manageable steps for better performance. The system utilizes three core functions: The Perception layer reads and extracts data from messy documents, including scans and various PDF formats. The Semantic/graph layer makes sense of extracted data and establishes how all pieces relate to each other. LLMs and agents operate on this structured, meaningful information to perform complex reasoning tasks.
What Did Buchner Explain?
Buchner explained that a door in a drawing is often represented by an arc on a wall. A trained eye learns to interpret this visual language. The perception layer teaches the AI to read this language, and the semantic layer gives it meaning. It connects the door to the drawing, the governing spec, and the installing trade. This allows engineers to ask, “Does this door create a problem down the line?” rather than just, “Is there a door here?” Conflict caught during the design phase is relatively low cost. A similar problem found during construction costs tens of thousands of dollars to fix.
Trunk Tools maintains that its agents achieve approximately 95% accuracy in processing complex documents. The company trains all three layers on very specific datasets provided by customers. Data de-identifies and collects across the platform. Trunk also gathers labeled data through other paths, such as 3D building information modeling (BIM). The team uses continuous evaluation pipelines. These pipelines use ground truth data from experts to keep the system performing well. They also employ an LLM-as-a-judge framework. This framework creates a composite score that tests a model’s performance or risk. Related context: AI coverage.
