Why Passing Every Test Isn't Enough: Tensorleap's Bet on Explainable AI

A car's safety system misses a baby seat because of one blind spot in its training data. Tensorleap's CEO built a company to catch failures like this before they reach the real world.

Every time a car passed through one automotive safety company's testing pipeline, its AI model would miss a child's car seat if it was covered by a blanket. Not occasionally. Every single time. The model had passed every standard test. It had a 99% accuracy score. Nobody caught the gap until Tensorleap's platform traced it back to a single, mundane cause: whoever collected the training data in the lab never happened to put a blanket over a car seat. The fix was straightforward once someone could see it. Finding it was the hard part.

That gap between "the model passed its tests" and "the model actually works" is the problem Tensorleap exists to solve.

A quiet failure, at scale

Most teams building AI for the physical world, including autonomous vehicles, robotics, semiconductor manufacturing, and medical devices, rely on a similar shortcut: collect a large batch of data, train a model, run it against a test set, and if the accuracy number clears a threshold, ship it. It's a reasonable process, and it's also quietly broken in two ways.

First, the data teams test on rarely resembles the world their model will actually meet. A self-driving fleet trained mostly on familiar Bay Area routes has no real way of knowing how it will behave the first time it drives through an unfamiliar city. Second, and less obviously, the failures that do show up in testing often look random when they aren't. Engineers see a handful of oddball misses, find no obvious pattern, and label them "sporadic." Then the model reaches production and fails at the exact same corner, in the exact same conditions, every time. It wasn't sporadic. Nobody had the tools to see the pattern until it was already costing money, or worse.

That's the core issue: deep learning models are a black box. They can tell you a decision was made. They generally can't tell you why, and when something goes wrong, most teams have no way to find the root cause short of manually combing through thousands of samples by eye, a process that can take months for a single class of failure.

From high-stakes research to a startup

Yotam Azriel, Tensorleap's CEO, encountered this problem long before he had a company to build around it. Before co-founding Tensorleap in 2019, Azriel spent roughly a decade working in machine learning, including a role as a lead machine learning developer at the Soreq Nuclear Research Center. It's the kind of environment where a model's failure isn't a bad customer review, it's a decision that has to be defensible, and where "the model was 99% accurate in testing" is nowhere close to a good enough answer.

He co-founded Tensorleap alongside David Ben David and Nir Ben David to build the tool that environment never had: a way to see inside a neural network, not just watch its output. The company launched publicly in 2022 with a $5.2 million seed round backed by Angular Ventures, Sozo Ventures, and Industry Ventures, and added an $8 million round from Vertex Ventures Israel in early 2025. As the company scaled from early debugging tool to production grade platform, leadership evolved with it: Azriel, originally the technical co-founder, stepped into the CEO role.

What the platform actually does

Tensorleap calls itself an "applied explainability" platform, a description that undersells what it actually means in practice: the company redesigns the entire pipeline a team uses to build and monitor a model, from training through production, around the question of why. Instead of a pass or fail test score, teams get a map of which specific patterns, environments, or edge cases a model struggles with, and a recommended fix, whether that's a data change or a model change.

The efficiency gains are the kind that get a CFO's attention as much as an engineer's. Customers have cut the volume of data needed to train and validate a model by up to 95%, going from a billion samples down to a few million, while cutting production failures by as much as 90%. Hexagon, one of Tensorleap's customers, achieved a 40% reduction in dataset size without sacrificing model accuracy, alongside faster debugging and clearer visibility into what to fix next.

Translated into stakes rather than percentages: a semiconductor fab that doesn't catch a defect for a week has to recall an entire batch of wafers, a mistake that can cost millions and happen roughly once a quarter at an average factory. An autonomous vehicle company that can't explain why its car braked unexpectedly can't safely scale its fleet. A healthcare AI that gets something wrong without anyone knowing why isn't a bug report, it's a patient sent down the wrong treatment path.

The wager everyone is already making

Azriel is candid that there's no version of this technology that eliminates risk entirely. AI is already being trusted with decisions that used to require a human in the room, and organizations are making that trade because the alternative, not using AI at all, has its own costs, including the diseases it might otherwise help cure. What Tensorleap is betting on is narrower and more practical: that the gap between "the model looks fine in testing" and "the model is actually safe in the world" is closable, and that closing it is the only way physical AI earns the trust it's asking for.

As Azriel puts it, the industry's focus is shifting from simply building models to making sure they're ready for real deployment, and that's the moment Tensorleap was built for.