I was twelve the first time I worked on a job site. My dad was a carpenter in a poor, rural part of Austria, so I grew up around the work and the challenges we all know: the coordination, the paperwork, the way missed details can cost a crew a day, a week, or a relationship with an owner that took years to build.
By the time I founded Trunk Tools in 2021, I'd spent my whole life in and around construction. We've since hired many people with the same story: people who've scrambled to prep for an OAC meeting, kept RFI logs, reviewed submittals, missed a change in a bulletin, and know in their bodies what the pressure of this work feels like.
Two things were clear to me from the start.
First, there is too much information on a project for any team to fully comprehend. One early customer was building a high-rise in Manhattan. When their documentation synced into our platform, there were 120,000 files averaging 30 pages each. Three and a half million pages for one building, a stack three times the height of the structure itself. Buried in that volume are the discrepancies that upend schedules: coordination problems, spec conflicts, drawing details that quietly contradict each other.
Second, the workforce meant to catch those risks is aging faster than it can be replaced. Nearly 40% of skilled trades workers are already over 45, and an estimated 17 million construction workers will leave the trade over the next decade. This isn't a headcount problem you fix with a hiring cycle. It's an experience cliff: the judgment of a generation leaving faster than it can be passed down.
Customers kept asking us whether AI could actually help with the volume, the discrepancies, the buried risk, and the labor force transformation? And most importantly, could Trunk Tools make AI able to read drawings? Today, our answer to those questions is an emphatic "yes" thanks to Cortex, our new construction-specific intelligence layer that powers every Trunk Tools agent.
What We Built First
We started with TrunkText: a way for anyone on a project — in the field, on their phone, without walking back to the trailer — to ask a question and get a cited, context-aware answer in under a minute.
That sounds simple. It wasn't. Construction documents aren't isolated text files; a single wall section has to be read against three other sheets, a spec section, and a detail drawn six months earlier. We couldn't point a general-purpose model at that and call it done. Some early customers had already tried. So we built our first version of a construction-specific AI "brain" trained on construction data: the institutional memory of a 30-year superintendent who has read every sheet, spec, and RFI on your project and never forgets a thing.
Why We Had to Build Cortex
Because TrunkText worked, customers pulled us deeper: "Why can't it do this?" "It would be great if it could…"
So we turned next to one of the most time-intensive tasks in construction: submittal reviews. Reviewing a submittal properly requires awareness of all relevant spec sections and RFI responses that amend them, often with the tags missing in the system of record. Asking an early-career PE to review hundreds of submittals and catch every compliance gap invites rework and schedule delays that erode margin. We tested every generic model on this. None produced what was needed. So we built TrunkSubmittal, training our own models to infer the relationships between spec, submittal, and RFI automatically. One year later, the results are clear:
- 35,000 submittal reviews run on our platform — 60% non-compliant, 35% partially compliant, only 5% fully compliant, proving how much PEs and APMs are up against.
- Median run time of 3.92 minutes, against the 45–60 a proper manual review takes.
- Median cycle time down from 54.8 days at launch to 14.0 today — a 74% reduction.
- "Stuck" submittals (over 60 days for approval) down from 42.8% to about 2%.
The proof shows up in the field. A PE at Gilbane told us the architect on her Milwaukee project praised her attention to detail at an OAC meeting. An APM at Cleveland Construction said his architect recommended them to other owners in the area on the strength of their submittal process.
But drawing review is where the problem gets especially hard. It's where the gap between general-purpose and purpose-built AI matters most. Our platform data tells us that 10 to 20% of changes in a drawing set aren't clouded, and design teams include a quality narrative less than 5% of the time. How do you get in front of a problem you don't know is there?
Any PE who has fed a drawing set into a general-purpose AI model knows how short it falls. The research confirms it: frontier models, the same ones underlying most "AI for construction," top out around 26% accuracy on something as basic as detecting a door. Not because the prompts are wrong, but because drawings communicate through visual symbols no general model was trained to read. Better prompting doesn't close that gap. Different training does.
So we built it. There's no ImageNet for construction drawings, so we created the dataset ourselves, labeling symbols by hand, then scaling into a corpus of over two million labeled drawing artifacts. On that same door-detection task, we hit 97% precision and 90% recall. Cortex doesn't just read a sheet, it sees what changed when a bulletin arrives, including the changes the architect didn't cloud. For our beta customers, a 20-sheet bulletin that took up to six hours now takes under ten minutes, with written narratives, flagged risks, and RFI drafts ready to send.
Where We're Going
The momentum compounds as the distance between workflows disappears. TrunkRFI prevents unnecessary RFIs, drafts the necessary ones, and reads the implications of a design team's response. TrunkBid turns subcontractor proposals into an apples-to-apples leveling analysis for buyout. Visual Browsing lets teams pull every related detail, spec, and document by clicking a single element on a drawing. And the agents now hand off to each other: when TrunkReview flags a change worth a question, TrunkRFI drafts the RFI with the context already filled in. No re-entry, no dropped detail.
Cortex is what makes this possible. Underneath every agent is a construction ontology: a structured model of what a submittal, a spec section, an RFI, a drawing detail, and a schedule activity actually are, and how they depend on one another. On top of it, Cortex builds a knowledge graph that links every document and datapoint on your project: this spec section governs that submittal, this RFI amends that spec, this detail appears on these sheets and drives that schedule activity. The result isn't a system that retrieves text. It's one that understands relationships so every agent reasons with the full picture of your project, not one document in isolation.
Let me be clear about what this is not. It's not a general-purpose model pointed at construction. Not AI acquired last quarter and bolted onto a legacy tool. Not a toolkit we hand you to build and maintain yourself (your PEs shouldn't have to become prompt engineers to leverage AI to review a submittal). Our agents arrive pre-built and pre-trained, working the day your documents land, not a roadmap promise you'll see in two years. We price them as a predictable, up-front commitment, not a meter that runs out of credits the moment your team leans in. And we don't ship software and disappear. Our people are in the field with yours, because adopting AI is as much a change-management challenge as a technical one. We build, maintain, and run the agents so your team can stay on the build.
My conviction is firm. Unless you have a construction-specific intelligence layer, and unless every agent is measured against a ground-truth evaluation dataset, you are no better than what a general model produces on a good day. In construction, where every minute carries weight, that isn't good enough.
In the end, the technology only matters if it delivers. Builders on Trunk Tools catch risk before it compounds and save weeks of time. Their project teams start the day with a manageable to-do list because the tedious work is already done, and sleep better the night before an OAC meeting because we've taken many of the unknowns off the table.
That's why we build what we build. Cortex is the next evolution of our mission: to let builders build.