SixSense

This is less glamorous than most AI stories and more useful than most of them, too. The chips that run everything are manufactured in fabs - fabrication plants - where the margin between a good wafer and scrap is measured in particles you cannot see. Every step of the process generates inspection images, and traditionally those images get triaged by trained engineers or by rigid, rules-based classifiers baked into the equipment. Both approaches share a weakness: they do not really learn.

SixSense, founded in Singapore in 2018 by Akanksha Jagwani and Avni Agarwal, builds software that does. It uses deep learning and computer vision to detect and classify defects automatically - what the industry calls AI-ADC, for automated defect classification - and then extends that into the decisions that actually cost money: which wafers are high-risk, what the root cause of a recurring defect is, and which lots to disposition how.

The clever part is not that SixSense trained a model. Everyone trains models. The clever part is who they let drive it. The platform is no-code, aimed squarely at the process engineer who knows the fab intimately and has never written a line of Python. That engineer can take the factory's own defect images, fine-tune a model, and deploy it in under two days - without a data-science team standing between them and the line.

If you have spent any time around enterprise AI, you know that the gap between "we have a model" and "the model is running in production and someone trusts it" is where most projects quietly die. SixSense's design choice is to collapse that gap. That is a product decision as much as a technical one, and it is arguably the reason customers like GlobalFoundries and JCET - not companies that buy hype - actually deploy it.

The numbers the company cites are the kind that get a CFO's attention rather than a Twitter thread's: up to 30% faster production cycles, a 1-2% improvement in yield, and a roughly 90% reduction in manual inspection work. In semiconductors, a couple points of yield is not a rounding error - it is the difference between a profitable line and a marginal one, multiplied across enormous volume. That is the math SixSense is really selling, and it explains why "boring" AI in a fab can be worth more than a flashier model somewhere else.

To date the platform has analyzed more than 100 million chips. That is both a marketing statistic and, more interestingly, a training-data moat: the more defects it sees across more devices and more manufacturers, the better it gets at recognizing the next one - including recurrence patterns that span different products. Defect knowledge, in other words, compounds.

There is also a timing story. Geopolitics is busy redrawing the map of where chips get made, pushing new fab investment across Southeast Asia and back into the United States. New fabs need inspection intelligence on day one, and a software layer travels a lot more easily than a building full of metrology tools. SixSense, headquartered in Singapore with a foot in San Francisco, Bengaluru and Hsinchu, is positioned to follow the fabs wherever policy sends them.

None of this requires you to believe SixSense will win. The semiconductor toolchain is dominated by giants - KLA, Applied Materials, Onto Innovation - and incumbents rarely cede the fab floor gracefully. But the wedge is real: a nimble, no-code AI layer that engineers can actually operate, sold on yield economics that are easy to verify. That is a defensible place to stand.