Trunk Tools cuts document review from 60 days to 10 with vertical-specific AI stack
Trunk Tools says it compressed construction document review cycles from 60 days to 10 by abandoning general-purpose models for a vertical-specific stack. Founder Sarah Buchner and CTO Amrish Kapoor describe a three-layer architecture that extracts symbols from messy PDFs and drawings, structures them into a knowledge graph, and then applies reasoning agents. This approach treats general-purpose LLMs as insufficient for high-stakes construction workflows where a 2-millimeter symbol variation can mean tens of thousands of dollars in rework costs.
The source frames Trunk's architecture as a response to the failure mode of foundation models in industrial verticals. Where general-purpose systems optimize for breadth, construction documents require precision on rare terminology and implicit spatial relationships that practitioners "just know." Buchner, a former carpenter, notes that construction drawings use symbolic notation where an unlabeled arc on a wall represents a door, not a geometric curve. The perception layer handles this extraction using computer vision techniques trained on construction-specific symbols, feeding a semantic layer that connects the door to its specification, governing trade, and downstream dependencies. This structure aims to solve what the company calls the "long-term project memory" problem, spanning months or years of documentation that exceed context window limits and probabilistic interpretation tolerances.
The technical stack relies on domain-specific training data and modular separation of concerns rather than monolithic model scale. Kapoor explains that all three layers train on customer datasets with explicit permissions, with opt-out options for sensitive data, supplemented by synthetic data from 3D building information modeling. The company maintains a 95% accuracy threshold before deploying agents, using continuous evaluation pipelines and LLM-as-a-judge frameworks for subjective scoring. This reliability focus addresses the specific economic structure of construction errors, where design-phase fixes cost relatively little compared to field-discovered conflicts that can trigger $10,000 rework bills or six-figure delays from missed fireplace sealing requirements. The source cites one case where an agent flagged $60,000 in unsubstantiated landscaping pricing and another where early identification of a fireplace sealing requirement prevented roughly $100,000 in labor and material costs.
Industry commentators cited in the source reinforce the strategic framing while noting the architectural tradeoffs. Senior product manager Kriti Faujdar observes that valuable enterprise data rarely enters pretraining corpora, leaving foundation models "weak at anything niche" despite being trained to be "okay at everything." She suggests pre-training on domain data followed by task-specific fine-tuning, proposing that mixture-of-experts architectures can provide specialization without proportional inference cost increases. Developer Sébastien De Bollivier advises hybrid stacks pairing general models for orchestration with fine-tuned smaller models or dense retrieval for domain extraction, noting that legal and healthcare share construction's combination of high error stakes and standardized document formats. Both commentators emphasize that retrieval-augmented generation alone merely supplies better facts to models that still cannot reason properly in the domain.
The measurable outcomes follow the architectural bet on vertical depth over horizontal generality. Trunk reports average time savings of 8 minutes for single-document retrieval, 20 minutes for standard cross-referencing across two to three specification sections, 40 minutes for multi-document research spanning four to six documents, and 75 minutes for complex tasks like RFI generation and change tracking. The company says its submittal review agent specifically reduced cycles from 50-60 days to 10 days by flagging missing, conflicting, or noncompliant information in product specifications. In one instance, an agent identified an 8.5-inch undocumented beam relocation that would have required stripping and reinstalling structural elements, avoiding what Buchner estimates as $10,000 or more in direct costs plus schedule impacts. The agents now communicate autonomously, with drawing review agents handing off to RFI agents when discrepancies require clarification.
The vertical specialization creates predictable constraints that limit portability. Faujdar notes that specialized models often fall apart outside their training domain, making them unsuitable for general use without retraining. The source also acknowledges latency risks, as increased reasoning capacity in underlying models can slow response times. Buchner states the company evaluates latency tradeoffs against user experience impacts before releasing updates, ensuring marginal performance gains justify any responsiveness degradation.
The construction-sector results validate that domain-specific preprocessing and knowledge graphs can compress review cycles dramatically when error costs are high and document formats are standardized. Whether this three-layer architecture transfers to other verticals depends on whether they share construction's combination of symbolic document density, proprietary data silos, and nonlinear error costs, none of which the source benchmarks across industries.