Somewhere on a production line right now, a server is rolling past a ring of cameras. No inspector leans in. No flashlight. No clipboard. A box of software the size of a hardback book looks at the assembly, decides whether it passes, and writes down why. The cameras belong to the factory. The judgment belongs to Relimetrics.
This is the ordinary scene the company exists to create: quality control that does not depend on a tired pair of human eyes at the end of a long shift. Relimetrics, Inc. builds ReliVision, an AI visual inspection platform that manufacturers install themselves, train themselves, and - this is the unusual part - trust enough to stake their reputation on.
"It has no fatigue-induced errors, and it moved performance from 2.1σ to 4.2σ."Senior Director, Supply Chain - Hewlett Packard Enterprise
The Problem They SawFactories that cannot see their own mistakes
Modern manufacturing has a quiet, expensive secret. The machines that build things are precise to the micron. The step where someone checks whether the thing was built correctly is often a person, squinting, on hour seven. Defects slip through. Good parts get scrapped. And nobody can fully explain either outcome after the fact.
The obvious fix - computer vision - had been promised for a decade. The catch was that traditional machine-vision systems were rigid. Change the part, the lighting, or the camera angle, and the system needed a specialist to reprogram it. Deep-learning systems were more flexible but demanded data scientists most factories did not employ. Either way, the AI lived in a lab, not on the line.
The promise was always "AI for the factory." The reality was usually "a PhD for the factory." Relimetrics noticed the difference.
The Founders' BetA fracture-mechanics PhD walks onto a shop floor
Relimetrics was founded in 2015 by Kemal Levi, who earned a doctorate at Stanford in fracture mechanics and non-destructive testing - the science of finding flaws in things without breaking them - and an MBA at UCLA. He was joined by co-founder Burak Acar, PhD. The pairing is telling: one half steeped in how materials fail, the other in how to teach software to recognize patterns.
Their bet was contrarian. Instead of selling AI as a service that the manufacturer would forever depend on them to operate, they would hand the controls to the customer. The engineer who knows the part best - not a remote data scientist - should be the one teaching the model what a defect looks like. The software would be hardware-agnostic, working with whatever cameras a plant already owned, and it would run on the edge or in the cloud as needed.
It was a harder product to build. It was also, they wagered, the only version manufacturers would actually keep using. The company set up two homes - one in Silicon Valley, one in Berlin - which is either ambitious or slightly stubborn for a team that has stayed deliberately small.
ReliVision, explained without the jargon
ReliVision does four things that, together, are the whole pitch. It lets a team curate and annotate images with a graphical tool instead of code. It ships with a library of pre-trained models so deployment starts in days, not quarters. It allows single-click retraining, so when a new defect appears, the operator who spotted it can fold it into the model. And it generates synthetic image variations - realistic fakes that reflect real lighting and wear - so the model learns from edge cases that rarely show up in time.
Retrain the model with one click. Try negotiating that with a night-shift inspector.
The architecture is built to spread. One camera becomes one line becomes a dozen sites, all synchronized. Operators on the floor send feedback through a plain web interface, which means the system gets smarter from the people closest to the product rather than from a distant queue of labeling contractors. The applications stretch from cosmetic surface defects and automotive module inspection to wind-turbine blades, X-ray and ultrasound non-destructive testing, and construction panels.
The hardware-agnostic part is easy to skip past and worth lingering on. Most inspection vendors prefer to sell you their cameras, their lights, their box - a tidy bundle that is also a lock. Relimetrics took the opposite tack: bring your own sensors. That choice makes the first install cheaper and the eventual rip-out harder, because the value lives in the model the customer trained, not in a rack of proprietary metal. It is a quieter form of stickiness, earned rather than imposed.
The Relimetrics File
When the insurer signs the homework
Anyone can claim accuracy numbers. What is harder to fake is someone else putting money behind them. In 2022, the reinsurance giant Munich Re brought ReliVision into its aiSure program, which guarantees the detection accuracy and shop-floor availability of the software. Before underwriting it, Munich Re ran its own technical due diligence. That is the part worth pausing on: a company whose entire business is pricing risk decided this AI was a risk it could price.
A reinsurer will not insure a coin flip. It insured ReliVision.On Munich Re's aiSure backing, 2022
The customer roster reads like a tour of things that must not fail. At Foxconn, ReliVision inspects complex server assemblies built for HPE - the deployment that produced the jump from 2.1 to 4.2 sigma and cut the defective products reaching customers. With Lockheed Martin Aeronautics, the company worked on AI-driven machine vision for robotic automation in aircraft manufacturing. With Siemens Gamesa, it tackled non-destructive testing of wind-turbine blades, where a missed flaw is a problem measured in megawatts and altitude.
Notice what those four industries have in common: aerospace, electronics, energy, construction panels. They are not the easy demos. They are the sectors where an escaped defect turns into a recall, a grounding, or a headline. A startup of roughly a dozen people does not usually get invited into rooms like those. Relimetrics did, which suggests the technical due diligence held up not once but repeatedly, in front of buyers with every reason to be skeptical.
From 2.1 to 4.2 sigma
Hand the AI to the people who know the part
Strip away the demos and the company's mission is almost modest: let manufacturers build, deploy and own their AI inspection without coding or an AI team. Relimetrics describes its staff as an interdisciplinary group of engineers and developers - "25+ AI Pioneers" - which is a small number for a footprint this wide. The smallness is part of the point. If the product needs an army to run, it has failed its own thesis.
Most AI companies sell you a dependency. Relimetrics sells you independence and then, oddly, hopes you do not need to call.
There is a tidy irony in a non-destructive-testing scientist building software whose success is measured by how little anyone has to touch it. The better ReliVision works, the more invisible it becomes - just a quiet box on a rail, doing the one job no human ever truly enjoyed.
Why It Matters TomorrowReliability is becoming a feature you can buy
As factories chase Industry 4.0 - more variants, shorter runs, higher complexity - the gap between what machines can build and what humans can reliably check only widens. The interesting frontier is not just detecting defects but proving you detected them, in a record clean enough that an insurer, a regulator, or a customer will accept it. Relimetrics is betting that verifiable, self-owned AI quality is about to become table stakes rather than a differentiator.
If they are right, the strangest thing about that server rolling past the cameras is that one day no one will find it strange at all.
The future of quality control is not a smarter inspector. It is the moment you stop needing to look.
Back to where we started. The server reaches the end of the line. The cameras have already decided. The book-sized box has already written down why. Five years ago that judgment would have waited on a person at hour seven. Now it is finished before the part stops moving - and somewhere, an insurer has agreed that the machine was right. That is the scene Relimetrics set out to build. On a growing number of factory floors, it already looks ordinary.