Behind the Pins: How We Built a Smarter Way to Inspect Connectors

The story behind the AI model that’s making 98% recall look easy.

What does it take to build a best-in-class AI model for one of the hardest inspection problems in electronics manufacturing?

Thousands of tiny pins. Dozens of red rabbits. A team of engineers armed with tweezers, glue, and no fear of getting their hands dirty.

This is the story of how we at Instrumental engineered not just an AI model, but a fully representative training dataset—one precise defect at a time.

Why Pins Are So Hard

High-density pin arrays—like the ones used in CPU and GPU trays—are a complete nightmare to inspect. Thousands of nearly identical, microscopic metal pins must be perfectly aligned. A tiny bend? That’s a failed test. A single strand of hair? That’s a connectivity issue. A bit of oxidized residue? That’s an RMA waiting to happen. But as engineers, you know this. 

Traditional AOI systems can’t handle this complexity. They're brittle, rule-based, and require hours of calibration. And even then, they miss the weird stuff. Manual inspection isn’t better—it’s slow, inconsistent, and tops out around 80% recall.

Our goal was to create a system that could inspect pins like a pro—and keep getting smarter over time.

The Turning Point: Ditching the Traditional Approach

When designing AI models for manufacturing, there are two main approaches:

• Simple Models: These require highly consistent imaging conditions to perform effectively, making them less adaptable in dynamic environments.
• Smart ML Models: Trained to handle variations naturally within the dataset, adapting to real-world inconsistencies.

At first, we did what everyone does: try to make the problem fit the tools. 

We’d been here before. At Instrumental, we build AI that helps manufacturers catch quality issues early—so we know how to approach inspection problems. 

Our team has built representative datasets to support building some of the world's most admired electronics.

But this problem was different.

We tried simple models with strict imaging requirements. We optimized lighting. We tried perfect alignment. The models still came up short. 

Eventually, we stopped fighting the variation and leaned into it. We turned our focus to what we do best: creating smart AI with representative datasets. That meant building a real-world library of real-world pin defects—no more synthetic data, no more idealized test units.

And that’s when things got interesting.

Inducing Defects (a.k.a. Red Rabbit Season)

Our solutions team using a lighter to create defects.

If you want your AI to catch defects that happen in the real world, it needs to see real-world defects. And that means you have to create them.

Creating our dataset meant we had to become pin defect artists. Bent pins, missing pins, smudged pins, FOD, burns, dents—you name it, we had to recreate it.

We spent hours with tweezers, microscopes, and a surprising number of makeshift tools inducing these issues. 

Was it tedious? Yes. Was it fun? Also yes.

We debated what counts as “realistic burn damage.” We argued over whether a tiny wire was “too clean” to be true FOD. We even tested different contaminants (including actual hair) to mimic real line conditions.

Smarter Setup, Better Optics

To detect microscopic issues, you need crystal-clear imaging. But telecentric lenses weren’t cutting it. We worked with optical engineers to develop a better setup—custom optics and lighting that minimized shadows and reflections and made every pin pop in high-resolution clarity.

From there, our AI kicked in: Synchronized Learning allowed us to start detecting defects from the very first unit. No golden samples. No long setup cycles. Just results.

In our validation run, we trained the model on only five connectors—each with about six defects—then tested it on 25 more. It crushed it: 94% recall, and ultimately 99% as the system learned and refined itself.

Why It Works: Train on One, Learn from Thousands

Here’s what makes our approach different: the model doesn’t just learn from one unit. It learns from thousands of image data points on that unit. That means our very first unit is essentially a crowd of training data, and the model can start surfacing meaningful insights on day one.

Better yet? It gets smarter with every unit it sees—and improvements made in one factory line instantly benefit all others.

What This Means for the Line

For our customers—many of them building AI infrastructure at scale—pin inspection is the difference between hitting throughput targets and missing them. Our system means fewer escapes, less rework, and no surprise RMAs caused by hidden connector issues.

It also means faster ramps, easier setup across new SKUs, and more confidence in every tray that leaves the line. You can get the full white paper on how Instrumental Synchronized learnings performs in pin inspection here. 

Final Thoughts from the Floor

This model didn’t come out of a lab. It came from the floor—from painstaking hours replicating defects, arguing over bent pins, and sweating the little stuff so the AI wouldn’t have to.

It’s one of our proudest builds. 

Want to talk to an engineer who built this? Or talk to an engineer about how this works in your factory? Chat with us.