Trunk Tools Accelerates Document Review Process from 60 to 10 Days by Transitioning Away from General-Purpose Models

In many sectors, the operations are far from the streamlined SaaS databases that one might expect; instead, they often consist of messy documents, proprietary formats, implicit workflows, and extended tasks, making it challenging for general models to handle them effectively.

This scenario led Trunk Tools, a company specializing in construction project management, to develop a unique three-layer system consisting of perception, semantics, and agents, designed to manage detailed data and enhance industry automation with high relevance and accuracy.

According to Trunk, their tailored technology has significantly reduced review timelines from several months to mere days, eliminated costly mistakes in the field, and enabled autonomous agents to analyze vast amounts of documentation accurately.

Trunk’s founder and CEO, Sarah Buchner, a former carpenter, explained their mission: “We aimed to consolidate data from various systems, process it, structure it, build an ontology into a knowledge graph, and subsequently train AI models.”

This innovative method could serve as a model for organizations in various fields looking to convert data disarray into automated, industry-specific processes.

Challenges with General-Purpose LLMs in Specialized Industries

While foundational language models (LLMs) are powerful, they excel in broad applications rather than specialized niches. Kriti Faujdar, a senior product manager focused on AI infrastructure, noted, “General-purpose LLMs tend to be average at everything, which makes them ineffective in niche areas.” This includes handling rare terminology, domain-specific reasoning, and implicit knowledge that experts naturally possess.

Web and software developer Sébastien De Bollivier shared concerns about the limitations of these models, stating that while advanced models like GPT-4 can interpret general legal documents, they struggle with specific references necessary for experts.

Moreover, Faujdar pointed out that valuable enterprise data often resides within internal systems, away from pre-training datasets. “RAG offers some improvement,” she said, “but it merely enhances the facts without solving the reasoning issues inherent to specific domains.”

Fine-tuning on domain-specific data after pre-training is crucial, as real-world examples are far more effective than numerous noisy data points. Techniques like mixture-of-experts (MoE) provide specialization without inflating inference costs. Combining RAG with fine-tuning can also yield enhanced results by improving factual accuracy while addressing vocabulary and reasoning flaws.

De Bollivier highlighted the benefits of hybrid approaches, using a general model for reasoning and orchestration, along with a finely-tuned model for domain-specific extraction. He emphasized that fine-tuning should focus on enhancing reliability with respect to specific output formats, rather than just making the model more domain-smart.

Industries like construction, legal, and healthcare are actively benefiting from these advanced techniques, given their high error stakes and standardized documentation formats, which yield measurable training returns.

Faujdar issued a noteworthy caution: specialized models often struggle outside their domain expertise and may require retraining to adapt.

Trunk’s Structured Three-Layer Stack

In highly specialized areas such as construction, a simple data dump into a large language model is insufficient, asserted Trunk’s CTO, Amrish Kapoor. This limitation arises because most transformers are probabilistic models, offering uncertain interpretations of inputs.

In construction, where minute symbols can have vastly different meanings depending on context, precision is paramount. There’s also the challenge of short-term context limits, as these models cannot manage long-term project memory that stretches over months or years.

To address these issues, Trunk has structured its workflow into three distinct layers:

• Perception: Extracting and reading data from chaotic documents, including PDFs and drawings.

• Semantic/graph layer: Providing context and understanding the relationships within the data.

• LLMs and agents: Utilizing AI on top of the structured data.

According to Buchner, construction drawings frequently employ symbols that may not clearly articulate their meanings. The perception layer trains AI to decipher this specialized language, while the semantic layer connects these details to related specifications and tasks. This functionality allows project engineers to ask insightful questions, addressing potential issues far earlier in the workflow.

The Scale of Unstructured Data in Construction

Buchner remarked that construction workflows are filled with implicit assumptions and intricate data connections, making processing unstructured information practically impossible. She estimates that the documentation for an average high-rise project can reach about 3.6 million pages—if stacked, it would rival the building’s height.

Trunk’s three-layer system is trained using particular datasets with explicit customer permissions. Customers have the option to withdraw their data from training purposes. The data is anonymized and aggregated, supplemented by additional labeled data sourced from 3D Building Information Modeling (BIM).

The company ensures that only agents with around 95% accuracy are deployed. Continuous evaluation is conducted based on data from customers and experts, along with a novel approach of using LLMs to assess performance objectively and subjectively.

Nonetheless, Buchner noted that increasing the reasoning capabilities of models could introduce latency issues. Trunk maintains strict criteria for evaluating latency whenever changes are made to their infrastructure prior to release.

Impressive Time Savings with Trunk’s AI Agents

Trunk’s AI platform features seven agents specifically designed for construction tasks, including analyzing requests for information, reviewing bids, and examining drawings. The submittal agent, for example, quickly identifies any gaps or inconsistencies in product specifications.

This agent considerably speeds up submittal cycles, cutting the time required from 50-60 days down to just 10, positively impacting both schedules and budgets.

Trunk’s agents are now beginning to interact autonomously, leading to more efficient workflows. For example, one agent can verify the accuracy of architectural drawings and subsequently pass that information along to other agents for follow-up questions.

Users report significant time savings, such as:

• 8 minutes saved on single-document retrieval.

• 20 minutes saved on standard referencing.

• 40 minutes saved on multi-document research.

• 75 minutes saved on complex tasks.

One noteworthy instance involved Trunk’s drawing review agent detecting an undocumented shift of a structural beam by 8.5 inches, which, if unnoticed, could have resulted in additional costs of over $10,000 for rework and significant delays.

Other alerts included identifying a $60,000 price inflation lacking justification from subcontractors and flagging essential preemptive tasks to prevent costly labor and material issues.

Lessons for Other Sectors

Trunk’s methodology for developing specialized agents can be applied across any sector managing substantial amounts of unstructured, industry-specific information. To maximize efficiency, businesses must first understand their specific data challenges and then establish a technical framework that allows for effective data management.

As significant investments flow into foundational models, organizations are encouraged to build modular systems that capitalize on the strengths of various models while also enhancing areas where generic models are lacking.