The client Stevens Industries is the largest manufacturer of commercial casework and architectural millwork in the United States. Despite operating a high-tech factory equipped with millions of dollars in advanced machinery, Stevens relied on a small team to manually review architectural drawings, extract requirements, and prepare project quotes. This labor-intensive process consumed hundreds of hours, making the ability to scale one that required additional skilled and hard to find labor. Here, we continue from the discovery phase already completed with Stevens, where we validated the project’s feasibility and identified a clear technical path forward. The Business Goal Stevens wanted to bring their estimation process up to date by automating the manual steps involved in scoping and quoting projects. Their goals were to cut down the time spent on estimating by at least 20%, reduce the chances of human error when reading complex architectural drawings, and free up their team to focus on higher-value work instead of repeating the same tasks over and over. Manufacturing Estimation, Part 2 In the first case, we explored how to approach automation in complex industrial environments, starting with a discovery process tailored to architectural blueprint analysis.
The Proof of Concept (PoC) phase was a key step for Stevens The main goal was to see if machine learning and AI could actually handle construction drawings well enough to deliver accurate estimates. Key goals of the PoC included: Technical Validation — The team aimed to determine how specific and granular the model could be in recognizing different types of cabinetry and related elements within the drawings. They outlined specific objects they wanted the model to identify, such as cabinets with varying features (e.g., doors versus drawers). Accuracy Assessment — The PoC focused on establishing the level of accuracy that could be achieved with the custom-trained model. This involved testing the model's ability to differentiate between various cabinet styles and configurations. Foundation for Development — The PoC served as a technical groundwork to confirm that the software architecture would support the intended functionalities. Once the PoC demonstrated success, it paved the way for moving into the Minimum Viable Product (MVP) stage, where the focus shifted to developing user interfaces and integrating features that would enhance the overall user experience. Overall, the PoC was essential in confirming the feasibility of the solution and setting the stage for further development and refinement.
PoC: 2 Key Stages As we described here, COXIT spent about 50 hours upfront verifying all the key assumptions: how big the blueprint files were, what made them tricky to analyze, which existing tools had already been tried, and whether any open-source computer vision models could offer a head start. Once the groundwork was in place, we launched a 160-hour sprint with Stevens’ estimators reviewing sample drawings, clarifying tricky cases, and labeling data. Together, we defined the first 14 objects for the AI to detect, like wall cabinets, base cabinets, and elevational markers from the floor plan.
Building the first working demo November 2023 The next step was to build a working demo that could read construction blueprints and identify the important parts. To do this, our team broke each giant PDF into smaller pieces so the AI could analyze them without running into memory issues. Then came the hard part: over 7,500 cabinets, notes, and other key elements were manually marked on a set of sample drawings to teach the model what to look for. After several rounds of training and fine-tuning, the results were optimistic. The model could now find 91% of the right elements and got things right 84% of the time. To make testing easy, we built a small webpage where the client team could upload a drawing and instantly get back a zip file with two things: – a marked-up version of the blueprint, and – a spreadsheet listing everything the AI had found.
Quick fixes December 2023 Early tests turned up a few issues. The model was missing some of the longest base cabinets, not because it couldn’t find them, but because the drawing had been split into tiles, and those cabinets got cut in half. The team updated how the system puts results back together. When it saw matching left and right pieces, it merged them into one. That fix made a big difference, and didn’t require retraining the model. There were a few smaller changes too. In busy drawings, the model sometimes flagged extra things that weren’t needed. Adjusting how overlaps were handled helped reduce those false positives.
On real projects the first model cut drawing review time from 2–6 hours (on average) to < 10 minutes, even though humans still gave the output a quick once-over. The success convinced Stevens to green-light full product development. [1] Object coverage — 14 essential elements like cabinets, doors, and construction notes, basically everything estimators look for on a drawing. [2] Seamless pipeline — From file to result in one flow: split the PDF, detect items, clean up the output, and export to CSV and PDF. [3] Easy access — No setup needed. Simply just drag, drop, and download from a web interface. [4] Built to grow — Train on new items any time by adding a few labeled examples. No retraining from scratch required.
Manufacturing Estimation, Part 3: Launching a Factory-Ready Estimation Tool for Stevens Industries Start reading
