Here is a fact about manufacturing that nobody puts on a brochure: every factory, no matter how good, occasionally ships a bad part. The industry even has a word for it - an "escape." A flaw slips past inspection, rides the conveyor to the loading dock, and becomes someone else's problem. Averroes.ai, a roughly fifteen-person startup in San Mateo, has built its entire company around the unglamorous, slightly obsessive goal of making that number zero.
The premise sounds almost too tidy. Manufacturers already inspect their products - often with expensive cameras, microscopes, and, in the semiconductor world, machines that cost more than a house. The problem was never the seeing. It was the deciding. Someone, usually a human, has to stare at thousands of images per shift and judge which speck is a harmless artifact and which is a catastrophic short circuit. Humans are good at this for about an hour, and then they are tired, and then things escape.
Averroes' pitch is that the deciding is a machine-learning problem, and that the machine can be handed to the factory rather than to a data-science team. The platform is no-code: an engineer uploads a modest pile of example images - the company says 20 to 40 per type of defect is enough - labels them, and the model learns what "bad" looks like. No Python, no GPU cluster procurement, no six-month integration. The claim is deployment to production in under five days, which, for anyone who has watched an industrial AI pilot die slowly over three quarters, is the number that actually raises eyebrows.
What makes the small-data claim interesting is that it inverts the usual story about AI. The popular narrative is that these systems are ravenous - that they need millions of labeled examples to do anything useful. But a defect is, in a sense, a rare event. A well-run line does not produce thousands of cracked wafers for the model to study. So a system that demands mountains of defect images is, by definition, useless to the customers who need it most. Averroes designed around the scarcity instead of complaining about it, leaning on techniques like active learning, where the model flags the images it is least sure about and asks a human to weigh in.
Catch every defect. Zero escapes.
That human is a feature, not a bug. Averroes describes its approach as "human-in-the-loop," which is a polite way of saying it did not build the robot that fires the inspector. The inspector stays. The model does the counting and the tedious first-pass triage; the expert validates the edge cases and, in doing so, teaches the model to be better next shift. Continuous learning, in the company's framing, means inspection stops being a thing you install once and becomes a system that adapts as your defects evolve - because they do evolve, as processes drift and new products come online.
The most quietly clever part of the business is what it does not ask you to buy. Averroes sells software, not hardware. It works with the inspection equipment a factory already owns, reading the images those machines already capture and simply interpreting them more aggressively. The company reports that customers catch 40 to 60 percent more submicron defects this way - not by installing better eyes, but by installing a better brain behind the existing ones. For a plant manager whose capital budget is already spoken for, "same cameras, more defects caught" is a rare and welcome sentence.
Then there is the reach. Averroes is deliberately vertical-agnostic, which is a bet that a defect is a defect whether it appears on a silicon wafer, a pharmaceutical tablet, a solar panel, a circuit board, a piece of packaged food, or the inside of an oil pipeline photographed by a drone. Skeptics will note that these are wildly different physical worlds with different tolerances and regulators. Averroes' counter is that the underlying task - look at an image, find the anomaly, classify it, measure it - rhymes across all of them, and that a single engine trained to generalize travels further than a stack of bespoke point solutions.
None of this happens in a vacuum. The AI-inspection space is crowded and getting more so, with rivals ranging from semiconductor-focused startups like SixSense to horizontal players like Landing AI and Instrumental, and incumbents such as Cognex and KLA that have owned the machine-vision aisle for decades. Averroes' differentiation is the combination: no-code, small-data, hardware-agnostic, and fast to deploy. Whether that bundle wins is, as always, a question of execution and trust - manufacturers are conservative for good reasons, and "trust me, ship it" is a hard sell on a line that runs 24 hours a day.
There is a nice symmetry in the name. Averroes - the Latinized name of Ibn Rushd, the 12th-century Andalusian polymath - spent his career arguing that observation and reason are not enemies. A modern inspection line is, if you squint, the same argument rendered in silicon: capture the world carefully, then reason about what you saw. The company would like you to notice that its founder, CEO Tareq Aljaber, spent 23 years shipping products to millions of users at Microsoft, Adobe, Atlassian, and Samsung before pointing that instinct at the factory floor. The lesson he apparently carried over is the least glamorous and most durable one in technology: adoption beats sophistication. A brilliant model nobody can deploy catches exactly zero defects.