Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / How Hardware Startups Use Optical Inspection to Catch Defects
Transcript
- Lucas: One of the hardest lessons a hardware startup learns is that a board can pass a functional test but still have a defect that will kill it six months later in the field. Luna: And that's where automated optical inspection — AOI — comes in. It catches things before they leave the line. Lucas: Exactly. AOI is basically a camera system that looks at every solder joint, every component placement, and compares it against a golden image or a set of rules. If a resistor is slightly rotated, or a capacitor is tombstoned — standing up on one end — the machine flags it. Luna: I've seen tombstoning in person. It happens during reflow when the solder paste on one pad wets faster than the other, and the component just tips up. On a dense board, you might not see it with the naked eye. Lucas: Right. And that's the whole pitch for AOI. A typical desktop system from a company like Keyence or Omron runs around five to fifteen thousand dollars. For a startup doing low-volume production — say a few hundred boards per batch — that's a meaningful but justifiable line item. Luna: Compare that to the cost of a field failure. I remember reading about a smart-lamp startup in 2025 that shipped two thousand units with a cold solder joint on the power regulator. They caught it during first-article AOI on the second batch, but the first batch was already in customers' hands. Lucas: And what did that cost them? I'm guessing a full recall plus replacements. Luna: Around a hundred thousand dollars, including shipping, refurbishment, and the hit to their Amazon rating. The AOI machine that would have caught it was eight thousand. Lucas: So the ROI is almost immediate if you have a single defect event. But let's be realistic — a lot of hardware founders I talk to skip AOI because they think it's only for high-volume manufacturing. They'll do visual inspection by a technician, maybe a magnifying lamp. Luna: The problem with human inspection is fatigue. After the first hundred boards, your eyes start to glaze over. AOI doesn't get tired. It also generates a log you can trace back. Lucas: That traceability is huge. If you do find a defect pattern, you can go back to your contract manufacturer and say, 'Look at the data from lot number 4 — we see a shift in placement accuracy on the QFN package.' Without AOI, that conversation is he said she said. Luna: There are two main types of AOI: 2D and 3D. 2D uses a single camera and measures colour, shape, and position. 3D adds a laser or structured light to measure solder joint height and volume. Which one should a startup prioritize? Lucas: If you're doing mostly passive components and standard QFPs, 2D is enough. But if you have ball-grid arrays or QFNs with hidden joints, 3D gives you that height profile. The catch is that 3D systems cost two to three times more. Luna: And speed. A 3D inspection takes longer per board. For a startup doing fifty boards a day, that's probably fine. But if you scale to five hundred, you might need an inline system that keeps up with the pick and place machine. Lucas: Inline AOI is another jump — we're talking fifty to a hundred thousand dollars. But many startups outsource that to their contract manufacturer. If you're using a CM that already has AOI on their line, you can often add it as a line item for pennies per board. Luna: That's a good point. You don't have to own the machine. You just have to specify AOI in your build instructions. And most reputable CMs have it. Lucas: But there's a nuance. The CM's AOI program is generic. They might use default thresholds that flag a lot of false calls — boards that look slightly different but are actually fine. If the false call rate is high, the operator might start ignoring the machine. Luna: So you need to tune the algorithm to your specific board. That means spending time on the first article with the CM, adjusting the lighting and the pass-fail boundaries. Lucas: Exactly. One startup I know — they make a wearable sensor for industrial workers — spent two full days at their CM's facility dialling in the AOI. They reduced false calls from thirty percent to under two percent. That made the line efficient enough that the CM gave them a small discount on later runs. Luna: That's a great example. The time investment upfront pays off in yield and relationship. Now, what about the algorithm side? Is AOI mostly rule-based these days, or are people using machine learning? Lucas: It's a mix. Traditional AOI uses pixel comparison and edge detection. But newer systems, especially the ones from startups themselves, are training neural networks on defect images. The advantage is that ML can generalise better — it can spot a novel defect that the rule-based system wouldn't flag. Luna: But ML also introduces the black-box problem. If the algorithm says a board is bad, you can't always explain why. For regulated industries — medical devices, aerospace — you need that traceability. Lucas: Right. So if you're building a Class II medical device, you probably want a rule-based AOI where every decision is attributable to a specific measurement. ML is great for consumer electronics where speed and pattern recognition matter more than documentation. Luna: Let's talk about lighting. That's the unsung hero of AOI. The angle, colour, and intensity of the lights determine whether a solder joint looks shiny and concave or dull and convex. Lucas: Most systems use ring lights with multiple colour LEDs — red, green, blue — at different angles. A good solder fillet will reflect light differently than a cold joint. The algorithm learns that signature. But if the lighting changes — say the LED ages or the board colour varies — you can get false calls. Luna: That's why regular calibration is important. Some startups forget to recalibrate between batches, especially if they switch board colours from green to black. Lucas: Black boards are notoriously hard to inspect because the contrast is lower. You might need to increase exposure time, which slows down the line. It's a trade-off. Luna: So what's the one piece of advice you'd give a hardware startup founder who's trying to decide if they need AOI? Lucas: I'd say start by analyzing your defect data. If you don't have any — because you haven't shipped yet — assume a defect rate of around one to two percent for a first run. Then estimate the cost of a field failure. If that number times your expected volume is greater than the cost of an AOI solution, buy it or pay your CM to run it. Luna: That's a simple expected-value calculation. And it's the same kind of thinking we use in other areas of business — it's just that in hardware, the cost of failure is physical, not digital. Lucas: Yeah, if you're building something that plugs into a wall or goes on a person's body, you have a moral obligation to test it. AOI is one of the most cost-effective ways to do that at scale. Luna: Speaking of building things that matter — we should mention that we can keep having these conversations because of listeners who support the show directly. We deliberately don't run ads on these episodes, and that's a choice we've made from day one. If you find value in the specific, no-fluff approach, there's a link to buy me a coffee dot com slash fexingo. It's a small way to keep this ad-free. Lucas: Yeah, and we mean it — no hidden agendas, no sponsor segments about mattresses you don't need. Just us digging into the details that help you build better hardware. So if that's worth a coffee, we appreciate it. Luna: Now, back to AOI — one thing we haven't touched on is the role of AOI after rework. When you hand-solder a repair, you should run that board through AOI again to make sure you didn't create a new defect. Lucas: Absolutely. Rework is a common source of latent defects. The heat from a soldering iron can damage adjacent components or lift a pad. AOI catches that. Luna: And with that, I think we've covered the essentials: what AOI is, when to use it, how to tune it, and how to pay for it. Next episode, we'll talk about something completely different — or maybe we'll go deeper into the ML side of inspection. Lucas: I'd love that. For now, if you're building hardware, get your hands on an AOI report from your CM. Look at the defect Pareto. That's where the learning starts.