The Real Opportunity Was Process Automation

Most people think about this problem as AI image recognition.

But the bigger challenge is business process automation.

The real workflow looks like this:

1. Import architectural drawing
2. Identify relevant plan views
3. Detect cabinets and appliances
4. Extract measurements
5. Convert to 3D layout
6. Export to CAD format

Traditionally, this process takes hours of manual work.

So instead of building a tool that just detects cabinets, we designed a complete automation pipeline.

Step 1: Converting Drawings into Machine-Readable Data

Architectural drawings are typically delivered as PDF files.

From a computer vision perspective, that’s a problem.

A PDF drawing is just a collection of:

• lines
• shapes
• numbers
• symbols

Everything looks identical to a machine.

So the first step in the system is simple but critical:

Convert every drawing page into a high-resolution image.

We process drawings at 300 DPI to preserve details that detection models rely on.

Step 2: Identifying the Right View

A single drawing sheet can contain multiple views:

• floor plans
• elevations
• section cuts
• cabinet details

Before we can extract cabinet data, the system must first determine:

Which view contains the actual layout.

We built a region detection step that segments the page into different visual areas and prioritizes the base floor plan for processing.

Without this step, the AI would attempt to interpret irrelevant sections of the drawing.

Step 3: Detecting Cabinets with Computer Vision

Once the base view is identified, we run a YOLO detection model trained on architectural drawings.

The model detects objects such as:

• upper cabinets
• base cabinets
• tall cabinets
• appliances

YOLO works well here because it provides fast detection with strong spatial accuracy, which is critical for layout reconstruction.

Each detection generates a bounding box with confidence scores, which helps filter unreliable detections.

Step 4: Extracting Measurements with AI

This is where things get complicated.

Cabinet dimensions appear in architectural drawings in formats like:

• 30″
• 2′-6″
• 34 ½”

They may appear at different angles and can be associated with objects through leader lines or proximity.

We built a measurement extraction pipeline using LLM vision models.

The process works like this:

1. Crop the drawing around each detected cabinet
2. Send that region to the model
3. Ask the model to extract width, height, and depth

The result is structured data attached to each cabinet object.

To maintain accuracy, we added validation rules.
If measurements fall outside realistic cabinet ranges, they are flagged for review instead of accepted automatically.

Step 5: Converting Detection Data into Real-World Coordinates

Object detection alone is not enough.

We also need real-world positioning.

Architectural drawings include scale references like:

1/8” = 1′-0″

The system identifies the scale marker and uses it to convert pixel distances into actual measurements.

This allows every cabinet to be positioned correctly in a coordinate system.

Step 6: Rendering the Layout in 3D

Once we have:

• cabinet types
• measurements
• coordinates

we can generate a 3D layout automatically.

We built a viewer using Three.js that renders cabinet structures as 3D objects.

This is where the system becomes useful.

Instead of building the layout manually, architects now:

• review the generated model
• move elements if needed
• correct detection errors
• add missing cabinets

They are editing the layout, not creating it from scratch.

That’s a major productivity improvement.

Step 7: Exporting the Layout to AutoCAD

The final step is converting the generated layout into AutoCAD DWG format.

Using the AutoCAD SDK, the system exports:

• cabinet geometry
• dimensions
• real-world coordinates

Once exported, designers can open the file directly inside AutoCAD and continue their workflow.

At this point, the automation pipeline has successfully converted a PDF drawing into a usable CAD file.

Where the System Stands Today

The current system is an MVP in active development.

It performs well on:

• clean digital drawings
• well-formatted plans
• modern architectural layouts

Edge cases still require improvement, especially:

• scanned drawings
• overlapping annotations
• older drafting styles

But even at current accuracy levels, the system already delivers a key benefit:

Architects correct the AI output instead of recreating the design manually.

The Bigger Lesson About Automation

Many automation projects fail because they focus only on AI models.

But real-world systems require something bigger:

Process design.

In this case, success required combining:

• document processing
• computer vision
• LLM interpretation
• geometry reconstruction
• CAD integration

Individually, none of these steps solve the problem.

Together, they create a complete automated workflow.

That’s the difference between building a model and building a production system.