Machine Learning-Powered Estimation Software for Stevens Industries

The client

Stevens Industries is the largest manufacturer of commercial casework and architectural millwork in the United States.

The goal

Their vision was to implement automation to help replicate the skillset and duties being performed by their current estimation team with a goal of at least a 20% reduction in time to complete the task.

The outcome

For Stevens’ staff, prior to creating the ML model, it took 2 to 6 hours to complete the review process. With the ML model in place, the review process now takes less than 10 minutes.

the problem

You see a 20-story office building. How many cabinets are in it?

Stevens Industries had built its reputation on precision manufacturing of casework and millwork for hospitals, schools, and commercial buildings in the US.

However, within their high-tech operations, one problem remained: manual project quotation.

stevens stevens stevens stevens stevens stevens stevens stevens stevens

The “Before” Process

Time-consuming estimation process created a potential barrier for growth

Use third-party software to find projects to quote
Manually review all project documentation
Before 2-6 hours
Determine if project aligns to Stevens services and products
Perform manual takeoff from architectural drawings of project for exact requirements
Build quotation to meet project requirements

It was ironic that despite millions of dollars of state-of-the-art equipment humming in the 500-employee factory, a huge opportunity existed due to the handful of people spending hundreds of hours reviewing complex architectural drawings, identifying project requirements, and producing a quotation to meet those requirements.

Delays and errors with complex, large-format blueprints, threatened the company’s growth plans and slowed down customer response times.

In their mind, Stevens saw a need to reduce the amount of time needed for lead processing. Their vision was to implement automation to help replicate the skillset and duties being performed by their current estimation team with a goal of at least a 20% reduction in time to complete the task.

The next step was to find the right partner.

Scheme Scheme

the solution

Can we use Machine Learning for that?

For this foray into the machine learning and object recognition, Stevens reached out to COXIT, a boutique AI/ML agency with a strong backend expertise.

COXIT had to make the project from scratch. Providing Stevens with a generic tool was not a viable option, especially sinceno suitable tool for such a specific task existed in open-source libraries at the time.

//Curtis Garrard,

Continuous Improvement Manager, Stevens Industries, Inc

We needed them to take an existing, manual process and design software automation to help replicate the current business process at a high speed and level of error proofing. Scope involved

  • manipulating large PDF based files
  • deploying a custom model library for object detection and recognition
  • backend data architecture to manipulate and reconnect data to drive value back to the end user

//Curtis Garrard,

Continuous Improvement Manager, Stevens Industries, Inc

Pre-discovery

For weeks, Iryna, the founder and engineer mastermind of COXIT, was running immersive studies and experiments to make sure COXIT could take on this project, and deliver as expected.

Before signing the contract, Iryna and the team spent 50+ hours on the evaluation and research phase, pre-testing and double-checking everything, from the available CV solutions to studying specifics of the manufacturing industry paperwork and Stevens’ software.

When the project was finally deemed viable, COXIT’s team kicked it off with a 3-hour work session alongside Stevens' experts to understand every nuance of the quote generation workflow.

With a comprehensive understanding of the challenge, the technical work has begun.

CTO Iryna spent dozens of hours to do pre-sale work and ensure the project can be fulfilled as imagined.

Discovery

//Curtis Garrard,

Continuous Improvement Manager,
Stevens Industries, Inc

They have done a great job of communicating both the high-level details (timeline, budget, and milestones) as well as the technical details (next sprint topics, current roadblocks, creative solutions).

//Curtis Garrard,

Continuous Improvement Manager,
Stevens Industries, Inc

First results

Discovery and first results

One hurdle emerged immediately: the sheer size and complexity of the architectural drawings. Direct processing of these images would overwhelm conventional machine-learning models.

COXIT employed a combination of image detection and pre-processing techniques. Giant drawings were cut into smaller pieces. These were then fed into the ML model, which was mathematically configured to focus on specific object recognition within the architecture drawings.

A sigh of relief echoed through the room when the system successfully identified the first complex object. It marked a pivotal shift.

chart stevens

The 160-hour discovery phase resulted in a working model with higher-than-expected recognition rates, right off the bat.

Discovery Phase Morel Performance

  • 91%

    of all objects on the drawings are successfully detected.

  • 84%

    of all detected objects are relevant (actually objects).

  • 87%

    of all detected objects are relevant (actually objects).

Though the first results came back strong, this was no time to relax. Graphs and accuracy scores are abstract concepts – what mattered was whether the model could see what human experts saw.

Besides, the discovery phase version worked only with 14 out of 151 objects. As the model was learning to recognize various types of cabinets and related products, the team was introducing it to more and more objects. The trick was to gradually increase the load in the next iterations while keeping the high-performance rates.

The ML team watched the first steps of the model and gradually fed it with knowledge and data.

stevens

POC stage and weeding out the mistakes 

There's a common misconception about ML models; people expect “the magical AI” to hit the ground running and give 100% precision right away.

The process, however, is more like teaching a child. There will be mistakes and funny misreadings before your student will make you proud with their performance.

For a few months after the discovery phase, COXIT and Stevens teams worked side-by-side to feed the model a dataset of real-world architecture drawings, marked with the precise object classifications and locations. New objects were added every day, and as with any project, it didn’t go without the bumps on the road.

At some point, the team noticed that the model doesn’t recognize one of the objects — the largest object on the list. The model had no problem counting the “mice”, but couldn’t see the “elephants”. The close investigation showed that this long cabinet got cut in the process of drawings segmentation, and only appeared to the model as fragments. Knowing this, the team was able to fix the problem.

the POC Phase Model Performance

  • 85%

    of all objects on the drawings are successfully detected.

  • 89%

    of all detected objects are relevant (actually objects).

  • 87%

    measure of accuracy, the harmonic mean if Recall and Precision, an F1 score

Step by step, the model learned to recognize 17 primary objects from the client’s list and do that with 89% accuracy, considered very high.

These 17 variations represent 95% of the object usually present in the drawings, which already lets Stevens automate almost all requirement gathering steps for their customer’s projects.

Scheme Scheme

The result after discovery and POC stages

Use third-party software to find projects to quote
Download project documentation into new software
2-6 hours10 min
Receive automated feedback if project aligns to Stevens services and products
take extracted data from architectural drawings and automatically transfer into quoting program
Finalize quotation to meet project requirements

Progressively, the ML model advanced to accurately identify 17 key objects from the client's inventory, achieving an 89% accuracy rate—considered exceptionally high. These objects account for 95% of the items typically found in the drawings, significantly streamlining the quotation process.

After PoC we discovered the way to partly automate PDF annotation process, and project quotation process, with discovered possibility of LLM integration.

Though the resulting output artifacts (PDF + CSV) still require a quick proofread by a person, at this point, the team is already saving a tremendous amount of time.

For Stevens’ staff, prior to creating the ML model, it took 2 to 6 hours to complete the review process. With the ML model in place, that same review process now takes less than 10 minutes.

Start of a great partnership

At the moment, the team keeps working on marking up more objects for the system to recognize automatically, adding features that’ll allow the model to tie the objects to SKUs from the Stevens database, and wrapping the solution into a user-friendly interface.

This project has become a win for Stevens Industries. Their investment in AI technology is yet to prove itself in the field, but the future looks bright.

A company that started in the 1950s in the garage now harnessed the power of artificial intellect. And with lead processing efficiency set to match their manufacturing capacities, exponential growth is in the stars for Stevens Industries.

//Curtis Garrard,

Continuous Improvement Manager,
Stevens Industries, Inc

COXIT provided a great partnership approach where they have started to feel like an extension of our company. With their ability to provide not only high-level support for complicated IT and computer automation projects but also their ability to understand the business value, we feel like anything is possible with their support.