Every AR-guided hardware experience starts with a 3D model. But the CAD files that engineers use to design products are very different from what you need for real-time AR on a phone. My job at VisionGuide is to bridge that gap — taking raw engineering files and turning them into interactive models that technicians can use in the field.

Here's what that process actually looks like, and why each step matters.

Why Engineering CAD Files Don't Work Directly

Engineering CAD models are built for precision manufacturing, not for real-time rendering on a mobile device. A typical CAD assembly for an industrial machine might contain:

Millions of polygons — every thread on every screw is modeled with geometric precision
Hidden internal geometry — parts that are inside other parts, invisible during normal use but still present in the file
Manufacturing-specific features — tooling paths, tolerance markers, draft angles that have no relevance for a technician
No semantic labeling — the CAD system knows this is "Part-2847-Rev-C" but doesn't know it's the "RAM slot" or "power connector"

Loading a raw CAD assembly on a phone would either crash the app or render at 2 frames per second. And even if it rendered perfectly, a technician couldn't interact with it meaningfully because nothing is labeled in human-readable terms.

The Preparation Pipeline

Step 1: Import and Assess

We receive CAD files in standard formats — STEP, IGES, Solidworks, Fusion 360 exports. The first thing I do is open the assembly and assess its complexity:

How many parts are in the assembly?
Which components will technicians need to interact with?
Which parts are purely structural and can be simplified?
Are there multiple configurations or variants?

A desktop printer might have 50-80 relevant parts. A medical X-ray machine might have 300+. An industrial CNC machine can exceed 1,000 parts. The complexity directly affects how much optimization is needed.

Step 2: Simplify and Optimize

This is where the most significant work happens. I reduce the polygon count while preserving the visual identity of each component:

Remove internal geometry that's never visible during repair or inspection
Simplify fasteners — a detailed screw model might have 5,000 polygons. A simplified version that looks identical from arm's length has 50
Merge static components that are never individually interacted with. The chassis frame doesn't need to be 40 separate parts if a technician never removes individual frame pieces
Optimize textures — replace high-resolution material textures with simpler alternatives that look good on a phone screen

The goal is to get the entire model under 50MB for smooth real-time rendering on mid-range Android devices. For complex machinery, that can mean reducing the total polygon count by 90-95% while maintaining visual fidelity at the interaction distance a technician would use.

Step 3: Label Every Interactive Component

This is the most critical step for AR guidance — and the one most people underestimate. Every component that a technician might need to identify, remove, replace, inspect, or connect needs a human-readable label with associated metadata:

Component name — not "Part-2847" but "Fuser Assembly" or "Primary Transfer Roller"
Part number — for ordering replacements
Interaction type — is this removed by pulling, unscrewing, sliding, or pressing a release tab?
Safety notes — "Caution: Hot surface" or "Disconnect power before handling"
Connection relationships — this cable connects to that port; this component sits inside that housing

For a medium-complexity device, labeling takes 1-2 full days. I work directly with the service team or product engineers to get the terminology right — using the exact words technicians use in the field, not engineering jargon.

Step 4: Define Interaction Zones

Beyond labeling, each component needs defined interaction zones — the areas where a technician's attention should be directed during a guided procedure:

Highlight zones — the outline that glows when the AR system is pointing at this component
Action zones — where exactly the technician needs to grip, press, or insert
Indicator positions — where labels, arrows, and instruction overlays appear without obscuring the work area

Getting these zones right is the difference between a helpful AR overlay and a confusing one. If the highlight zone for a cable connector is too large, it might overlap with adjacent connectors. If the instruction overlay appears right where the technician needs to see, it blocks their view.

Step 5: Test Across Conditions

The final step before a model goes into production is testing under real conditions:

Different lighting — the model needs to look correct under fluorescent office lights, dim server room lighting, and outdoor sunlight
Different angles — technicians won't always approach the machine from the same position
Different devices — the model must render smoothly on both a flagship phone and a 3-year-old budget Android tablet
Partial visibility — what happens when the machine is partially covered or in a tight space where only one side is visible?

I test each model on at least three different devices before signing off. Our app engineers Siva and Logesh have built the rendering pipeline to handle these variations, but the 3D model itself needs to be prepared correctly for the system to work reliably.

Common Challenges

Variant Management

A single product line might have 10-15 variants that look 90% identical but differ in specific components. Rather than creating separate models for each variant, we build a master model with variant-specific components that can be swapped programmatically. This keeps the total file size manageable while supporting the full product range.

Legacy Equipment

Newer products usually have CAD files available. Older equipment — sometimes the most common hardware in the field — often doesn't. For legacy equipment, we work from technical drawings, photographs, and sometimes physical measurements to reconstruct a 3D model. This is more time-intensive but essential for organizations that need AR guidance for their existing installed base, not just new products.

Balancing Detail and Performance

There's always a tension between visual detail and rendering performance. A beautifully detailed model that stutters on a technician's phone is worse than a simplified model that runs smoothly. My rule of thumb: if a technician standing at normal working distance can't tell the difference between the optimized model and the original, the optimization is good enough.

The Result

A well-prepared model enables everything downstream — accurate AR overlays, reliable component identification, meaningful step-by-step guidance. When the technician points their phone at the machine and sees the correct component highlighted with the right label and the right instruction, that's the result of the preparation work.

The model is only one part of the system — the workflow editor, the AI recognition, the mobile app all build on top of it. But without a properly prepared model, none of those layers can deliver accurate, trustworthy guidance.

How We Prepare CAD Models for AR: From Raw Files to Interactive Experiences