The client Stevens Industries is the largest manufacturer of commercial casework and architectural millwork in the United States. Despite having a high-tech factory with millions of dollars in equipment, a small team still spent hundreds of hours manually reviewing architectural drawings, identifying project requirements, and creating quotes. This process was slow and inefficient, creating a big opportunity for improvement. The Business Goal Stevens wanted to automate their manual estimation process to reduce project estimation time by at least 20%. Minimize human error in processing complex architectural blueprints. Improve scalability by freeing up their team’s time for higher-value tasks. The Discovery Goal Understand the problem in depth by identifying challenges, defining requirements, and reviewing existing processes. Explore whether an existing AI solution can help solve the problem before developing a custom approach. Plan the Proof of Concept (PoC), including tool selection, implementation strategy, and evaluation criteria
The discovery phase confirmed that machine learning could be used to automate the estimation process. Researched available datasets and existing Large Language Models. Selected a suitable LLM for further development. Conducted a 3-hour session with the client’s team to analyze their manual workflow, identifying time-consuming tasks and inefficiencies. Used these insights to create a structured plan for the Proof of Concept (PoC).
Discovery Phase: 3 Key Stages The Discovery phase focused on validating the technical feasibility and defining a clear implementation plan. It began with in-depth research led by COXIT team, to assess whether machine learning could effectively solve the problem
//Curtis Garrard, Continuous Improvement Manager, Stevens Industries, Inc LinkedIn 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
Technical Feasibility Study Iryna (LinkedIn) , the founder and engineer behind COXIT, led deep research and testing to ensure the project was viable. Iryna consulted with ML experts to assess whether the problem could be solved and how they would approach it. After confirming feasibility, an ML tech lead was brought in to research existing models and datasets. The team defined how data should be annotated, planned the Proof of Concept (PoC), and identified the necessary team for implementation.
//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).
Onboarding & Problem Understanding When the project was 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. This helped map out the full workflow and pinpoint key inefficiencies in the manual estimation process.
Next Steps: Detection model feasibility tests Going ahead, instead of building a full system prototype, COXIT ran targeted AI feasibility tests. They segmented blueprints into smaller parts to reduce complexity and improve precision. The key goal of the discovery phase was to understand how to approach a complex task that no off-the-shelf solution could solve. In situations like this, the research is an essential step to avoid wasting development hours on ineffective approaches.
Deliverables: A clear project mindmap, technical architecture, data processing plan, timeline, and budget — all outlining the scope, approach, and resources for next steps.
