Construction project management company Trunk Tools has built a specialized three-layer AI stack that slashes document review cycles from 60 days to 10, demonstrating how purpose-built architectures can overcome the limitations of general-purpose large language models when applied to complex, data-heavy industries. The company’s perception, semantics, and agents stack transforms messy, unstructured construction documentation — PDFs, drawings, scans, and proprietary schemas — into structured, agent-ready data that autonomous systems can reason over at scale. For organizations in any vertical wrestling with domain-specific data chaos, Trunk’s approach offers a replicable blueprint for building AI systems that actually deliver measurable, field-level results.
Why General-Purpose LLMs Struggle With Industry-Specific Data
Foundation models are optimized for breadth, not depth, and that tradeoff becomes a critical liability in specialized domains. “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 and agentic AI. Rare terms, domain-specific reasoning, and the unspoken context that any practitioner “just knows” fall outside the capabilities of models trained on general web data.
Software developer Sébastien De Bollivier identified the core bottleneck: reliability on data that is “jargon-dense, abbreviation-heavy, and format-specific.” He noted that a GPT-4-class model can understand a French legal contract but will fumble the specific article references practitioners need to cite. The most valuable enterprise data never made it into pretraining anyway — it sits in internal systems and proprietary formats. RAG helps at the margins, Faujdar said, “but it’s just giving better facts to a model that still can’t reason properly in the domain.”
The solution, according to Faujdar, involves pre-training on domain data, fine-tuning on high-quality task examples, and building custom evaluation pipelines. “A few thousand examples from real practitioners beats millions of scraped, noisy ones,” she said. Mixture-of-experts architectures can provide specialization without blowing up inference costs, and pairing RAG with fine-tuning works well: RAG handles the factual long tail while fine-tuning fixes vocabulary and reasoning. De Bollivier recommended hybrid stacks that pair a general-purpose model for reasoning and orchestration with a smaller fine-tuned model for domain-specific extraction. “Don’t fine-tune to make the model ‘smarter’ about a domain,” he advised. “Fine-tune to make it more reliable on the specific output format your workflow requires.”
Inside Trunk Tools’ Three-Layer Architecture: Perception, Semantics, Agents
In highly specialized domains, “data dumps” into LLMs don’t work, said Trunk CTO Amrish Kapoor. Most transformers are probabilistic models — they report that an image is “probably” a tree or “probably” a child playing next to a tree. For high-precision symbolic interpretation, that’s insufficient. In construction documents, a two-millimeter-wide symbol can have vastly different meanings depending on its placement on a page. Additionally, probabilistic models struggle with long-term project memory — not a context window of a few tokens, but memory that stretches across months and years, matching the duration of construction projects.
Trunk’s three-layer system breaks workflows into distinct stages:
- Perception layer: Reads and extracts data from messy documents like PDFs, drawings, and scans. Construction drawings are typically symbolic — a door isn’t always labeled “door.” Sometimes it is simply an arc on a wall that a trained eye learns to read over years. “The perception layer is what teaches AI to read that language,” said founder and CEO Sarah Buchner, a former carpenter.
- Semantic/graph layer: Makes sense of extracted data and understands their relationships. It connects the door to the drawing that details it, the spec that governs it, and the trade that installs it. This enables project engineers to answer critical questions — not “is there a door here?” but “does this door create a problem down the line?”
- LLMs and agents on top: Reason over the structured data and execute autonomous workflows. An agent might review an architecture bulletin and produce a visual overlay comparing older and newer versions, flagging additions and removals, then generate written narratives describing changes in simple terms.
The system identifies document type and begins extracting information based on content — drawings, schedules, paragraph text. That data is then transformed and augmented in the platform, triggering agentic workflows like knowledge graph relationships and end-user processes. The shift from probabilistic guessing to structured reasoning matters because the cost of problems compounds with time in construction. “A conflict caught in design is relatively low cost to address,” Buchner said, “whereas the same problem caught in the field might cost tens of thousands of dollars.”
The Scale of Construction’s Data Problem
Construction workflows are “ripe with implicit assumptions and connections between data in its myriad of sources,” Buchner said. The amount of unstructured data is “humanly impossible” to process manually. She estimated that the average high-rise building generates about 3.6 million pages of corresponding documentation. “If you print it into a stack of papers it would be as high as the building itself.”
All three layers of Trunk’s stack are trained on very specific datasets from customers with explicit permissions and auto-labeling and intellectual property protections. Customers who don’t want Trunk training on their data can opt out. Data is deidentified and aggregated, and Trunk also collects large volumes of labeled data through other pipelines like 3D building information modeling. The company only ships agents that achieve around 95% accuracy, maintained through continuous evaluation pipelines based on ground truth data from customers and experts. Trunk also employs an LLMs-as-a-judge model to score performance both objectively — a simple right or not right — and subjectively, where more nuance is required. For example, when creating an email or explanation, the LLM-as-a-judge framework produces a composite score that aggregates different metrics and tests model performance or risk.
Latency remains a challenge, Buchner noted; any time the reasoning capacity of underlying models increases, the risk of latency goes up. Trunk maintains evaluation criteria to objectively measure latency whenever changes are made to infrastructure, agents, or API calls. “Before we release to customers, we ensure marginal changes to the end-user experience are well worth the performance enhancements,” she said.
Measurable Results: From 60 Days to 10
Trunk’s platform powers seven AI agents purpose-built for construction, analyzing request for information responses, overviewing bids, and reviewing drawings and submittals. The submittal agent flags missing, conflicting, or noncompliant information in product specs and RFIs. Human reviewers previously had to compare documents across numerous other documents — a “super annoying workflow,” Buchner said. The agent completes the work in seconds, and Trunk reports it has reduced submittal cycles from 50 to 60 days down to 10, with massive schedule and financial implications.
Agents now communicate directly with each other. One agent will review an architectural drawing for accuracy, then autonomously hand it over to agents handling RFIs and asking follow-up questions. “If the drawings have problems, the RFI agent is taking over and is actively reaching out for clarification,” Buchner explained.
Customers report savings of 20 to 40 minutes per field question. Field users know the time suck of going back and forth from office trailers, digging through project documents in scattered systems or printed PDFs, reconciling discrepancies, and returning to coordinate with trade partners. Trunk reports these additional outcomes:
- 8 minutes saved per single-document retrieval (status checks, location lookups, quantity queries)
- 20 minutes saved per standard referencing task (cross-referencing two to three spec sections to form an answer)
- 40 minutes saved per multi-document research task (listing and filtering queries, mapping relationships, analyzing RFIs and submittals across four to six documents)
- 75 minutes saved per complex task (creating RFIs and other communication materials, deep cross-referencing across documents, change tracking)
In one instance, Trunk’s drawing review agent flagged that a structural beam had been moved up 8.5 inches — a change the architect had not documented. If it had gone unnoticed, the project manager would likely have had to strip out and reinstall the right size beam, adding $10,000 or more to the budget with schedule implications. Other examples include an agent flagging $60,000 in exaggerated pricing with no justification from landscaping subcontractors; identifying a fireplace that needed sealing before drywall installation, saving around $100,000 in labor, materials, and delays; and calling out that an electric door required a panel not included in electrical drawings.
Lessons for Any Industry With Unstructured Data
Trunk’s approach is broadly applicable to any vertical working with high volumes of unstructured, industry-specific data. Builders in specific verticals must understand the data challenges their end users face and build technical infrastructure that transforms unstructured data into something an LLM can traverse and understand, Buchner said. “Only then can you build the connections between data points that ultimately feed agentic workflows.”
With significant investment flowing into foundational models, enterprises should build modular systems that can leverage the strengths of various models as they continue to improve. The key advice: “Build your technical advantage where the generic models are not investing and not performing well.”
What This Means for Organizations Building Domain-Specific AI
Trunk Tools demonstrates that the most impactful AI deployments in industry verticals are not about plugging in the latest foundation model and hoping for the best. The real leverage comes from building a specialized pipeline — perception, semantics, and agents — that transforms domain-specific chaos into structured, queryable data. For AI and software leaders evaluating similar investments in legal, healthcare, or other document-heavy sectors, the takeaway is clear: invest in the middle layers of data structuring and ontology before layering on LLMs and agents. Start by auditing the specific document types, implicit workflows, and jargon that define your industry, and build the extraction and semantic infrastructure that turns that mess into a machine-readable knowledge graph. Only then will autonomous agents deliver the kind of cycle-time reductions, error prevention, and cost savings that Trunk is already achieving in construction.
