A Startup Built AI to Scan Racehorse Legs. Now It Inspects Oil Pipelines.

The strangest sales pitch in energy this year

A UK startup called ImpriMed VetVision spent three years building AI to find hairline fractures in horse legs. Then a pipeline operator in Texas signed a contract to use the same software to inspect 3,000 miles of oil and gas infrastructure.

The pivot happened because one person noticed something odd: to a computer vision model, a stress fracture in a racehorse tendon and a stress fracture in a steam pipe are both just grayscale pixels with a pattern.

How a vet conference became an enterprise sales event

The company's CTO was presenting at an equine medicine conference about automated tendon injury detection. One throwaway slide compared a horse leg ultrasound to a pipe weld x-ray.

A boiler inspector attending with his veterinarian spouse took a photo of that slide and posted it in a LinkedIn group for non-destructive testing engineers, asking: "Anyone else see what I'm seeing?"

Within a week, the startup had inbound messages from two pipeline companies, a chemical plant, and a shipyard.

According to the founder's later interview in Utility Analytics Weekly, they had no prior plan to sell into industrial inspections. They'd been trying to expand from large-animal clinics into small-animal vet practices, where AI imaging competition was already brutal.

Instead, they took a three-month detour to run a proof-of-concept with the Texas pipeline operator.

Why veterinary data works for industrial inspection

The startup had hundreds of thousands of labeled ultrasound and x-ray images of animal legs, tendons, hooves, and joints. Many showed hairline fractures or micro-tears as small as 0.3 to 0.5 millimeters — notoriously hard for human clinicians to spot.

A reliability engineer at a European district heating utility realized the pattern-recognition problem was nearly identical to finding early-stage cracks in welds and pipe bends.

So the company trained a new model variant on non-destructive testing images — ultrasound and radiographs of pipes and welds — but bootstrapped it using their much larger, better-labeled veterinary dataset. They only needed 20,000 labeled industrial images, a fraction of what a scratch-built model would normally require.

Because their cloud infrastructure, labeling tools, and workflow already existed, their incremental cost to serve the pilot was in the low tens of thousands of dollars.

The numbers that made it real

On a test set of 50,000 historic pipeline weld images, the AI flagged 92% of the defects that human inspectors had eventually found. The older QA system used by the operator caught 73 to 78%.

False positives — marking a good weld as bad — dropped by about 30%. That matters because every false positive means expensive re-inspection and potential shutdowns.

In live trials, the Texas operator estimated it could cut manual image review time by 40 to 50%, translating to a projected seven-figure annual saving in inspection labor and downtime.

On a pilot with a district heating network in Northern Europe, the AI found three critical cracks that two separate human reviewers had missed — in pipe segments already cleared for service. The utility's head of asset integrity said those three catches "paid for the entire pilot twice over."

What the cross-pollination means

The pilot converted to a multi-year SaaS contract that now accounts for roughly a quarter of the company's annual recurring revenue, according to a recent founder podcast appearance.

On the veterinary side, equine clinics using the software caught 30 to 40% more early-stage tendon injuries before they became catastrophic. Racehorse stables were willing to pay premium subscription fees for that edge.

The story challenges the conventional wisdom that AI startups must pick one vertical and stay there. This company did pick one vertical — and then discovered that the hard part wasn't domain expertise. It was building a workflow, labeling pipeline, and business model around any narrow imaging domain. Once that existed, new verticals with structurally similar data could be addressed with relatively little extra work.

The enterprise buyers essentially discovered a tool built for an unrelated industry and pulled it sideways into their own. Not through a formal adjacent-market strategy, but through a spouse at a conference, a photo in a LinkedIn group, and a Slack thread among pipeline engineers.

To the model, a fracture is a fracture. The context is just metadata.