A platform for engineers who can't fly to Asia at 2 a.m. to figure out why the new build is failing - and shouldn't have to.
Not a defect, exactly. Just a unit that looks slightly different than the 4,200 units before it. A screw seated half a turn too shallow. A gasket pinched by a hair. Last year, that unit would have shipped. This year, the line stops, an engineer in California gets a ping, and the failure mode is logged before the shift ends.
That is Instrumental, the quietest piece of infrastructure in modern electronics manufacturing. The company makes the software that connects the cameras, the test stations, and the messy spreadsheets - then trains a model to flag whatever does not look like the rest. Its customer list reads like a teardown of your desk: Meta, NVIDIA, Cisco, Bose, Axon, SolarEdge, ChargePoint, Motorola.
None of those companies build hardware to be looked at. They build it to be shipped. Instrumental is what they look at it with.
If you ship a phone, a server, a router, or a battery pack, you live inside a paradox. You design tests for every failure mode you can imagine. Then production starts and the failures that hurt you are the ones you did not imagine. A new vendor's adhesive. A subtly out-of-spec screw. A worker on the night shift who tightens a fixture a little differently than the day shift.
Traditional quality systems are built around known-knowns. They catch what the engineer wrote a test for. The unknown-unknowns - which is to say, most of them - leak through. Engineers find out about them three weeks later, in returns data, when the units are already in a customer's drawer.
The industry's old answer was to fly someone to Shenzhen. Instrumental's answer is to stop doing that.
Anna-Katrina Shedletsky and Samuel Weiss met at Apple, where Shedletsky was a product design engineer on the iPod and Apple Watch programs. The job involved exactly the kind of trips to Asia that Instrumental now exists to make optional. The premise of the company, which they founded in 2015, was almost embarrassingly simple: every modern assembly line is full of cameras. Almost none of them are looking. What if they were?
The technical bet underneath was less simple. Computer vision in 2015 needed enormous, labeled datasets - and high-mix electronics manufacturing has the opposite. Short runs. Constantly changing SKUs. A new revision every other month. So Instrumental built its system around discovery-driven anomaly detection: a neural network that learns what "normal" looks like from a handful of units - roughly 30 is enough to start - and then flags anything that drifts.
Instrumental's platform is not one product. It is three, and they only really make sense when they run together.
Vacuums up images, test results, and process data from every line, every shift, every facility. Stops the "where's that file?" portion of root-cause analysis.
Finds failure modes the engineering team didn't think to test for - the part of the platform that actually feels like science fiction the first time you see it.
Once a known issue is identified, deploys an automated inspection across every line in the world that builds the same SKU. No more "we fixed it in Vietnam, not yet in Mexico."
It is easy to say a platform "improves quality." It is harder to convince a Fortune 100 hardware org to publish a number. A few of Instrumental's customers have done it anyway.
Source: public statements from Meta and NVIDIA. Bars are relative, not normalized - which is honest, if inconvenient.
The shape of the customer list is the giveaway. These are not factories - they are brands. The contract manufacturers run the lines. Instrumental sits in the middle, owned by the brand, watching the manufacturer. That is a genuinely new posture in an industry where the brand usually has to take the manufacturer's word for what happened on the night shift.
Instrumental's stated mission is unfussy: eliminate the roughly two trillion dollars of waste sitting inside global electronics manufacturing - scrap, rework, returns, recalls, the engineering hours that go into firefighting instead of building the next thing.
It is, on its face, an absurd number. It is also, on closer reading, almost certainly an underestimate. The carbon attached to every piece of scrapped hardware is its own line item. So is the talent attached to every engineer flown to Asia to debug a problem a camera could have caught in an hour.
The mission is also, conveniently, the business model. Every defect caught earlier is revenue Instrumental can defend on the next renewal. Mission and metric line up - which, in a category as full of vague claims as "AI for manufacturing," is unusually grown-up.
The next decade of AI is, secretly, a hardware story. Racks, GPUs, switches, power supplies, the unglamorous physical layer of a trillion-dollar software boom. Every one of those boxes will be built on a high-mix electronics line. Every one will have defects somebody didn't anticipate. Every one will need to ship faster than the last.
This is the part of the Instrumental story that the company is not loud about, but the customer list already tells you. The platform that began catching cosmetic flaws on wearables is now in the supply chain of AI infrastructure. The boring quality-control problem turns out to be the bottleneck behind the most-funded category in technology.
So back to the line at 3 a.m. The camera notices something. The engineer in California gets a ping. The unit does not ship. The factory does not learn the lesson three weeks late, in returns data. Multiply that across Meta, NVIDIA, Cisco, Bose, every brand on that list, every shift, every night.
That is the company. Quietly, line by line, Instrumental is making the next billion devices a little less wrong on their way out the door.