ModelCat AI

The pitch is deceptively simple. Feed ModelCat your data, tell it which chip you plan to ship on, and its platform returns a machine-learning model tuned to that exact hardware - built, trained, optimized and validated - in something like zero to three days. The comparison it likes to draw is against the status quo, where getting a vision or audio model to run acceptably on a microcontroller can eat 12 to 24 months of specialist labor. That is not a rounding error. That is the difference between a product existing and not.

The word ModelCat uses for the machinery underneath is "AI-in-the-Loop." Stripped of branding, it means an AI agent orchestrates the whole workflow that a team of humans would otherwise slog through - a data scientist here, an embedded engineer there, someone who knows the dark art of quantization somewhere else. The agent picks the architecture, runs the training, squeezes the model down to fit, and checks that the result actually works. In one step, more or less, rather than a relay race of handoffs.

Here is the detail that separates the marketing from the engineering, and it is the detail worth dwelling on: ModelCat runs a hardware farm. Actual chips, wired up and powered on, so the platform can measure real latency, real memory use, and real power draw instead of estimating them from a datasheet. This matters more than it sounds. Edge AI is a game of tight constraints - kilobytes of memory, milliwatts of budget - and the gap between what a model should do on paper and what it does on a warm piece of silicon is exactly where projects go to die. Closing that loop with reality, not simulation, is the whole trick.

The company frames all of this against a market it sizes at $1.8 trillion, which is the sort of number that should be read with one eyebrow raised. But the underlying observation is sound, and its CEO puts it plainly: hardware capabilities are outpacing the software tools required to use them. Chips keep getting more capable. The tooling to actually deploy models onto them has not kept up. ModelCat is a bet that the tooling gap is where the value has migrated.

None of this appeared from nowhere. Eta Compute was founded in 2015 by Gopal Raghavan, Tim Semones and Paul Washkewicz, and spent its first act on silicon - most notably TENSAI, marketed as one of the lowest-energy AI microcontrollers around, aimed at devices that harvest their own power. It raised a $12.5 million Series C in December 2020, led by Synaptics. Somewhere along the way it built Aptos, a no-code cloud tool for edge machine learning, and in doing so seems to have discovered what it actually wanted to be.

The rebrand, then, reads less like a marketing exercise and more like a confession. The name is not subtle about its logic. Cats, the company says, are playful, quick-reacting, agile, and always land on their feet - which maps neatly onto a product that claims to pounce on a model build and deliver it with a near-zero error rate. Whether or not you buy the metaphor, it tells you what the thing is supposed to do.

The people running the current version of the company have the résumés you would expect. Evan Petridis, the CEO, is a serial entrepreneur who sold Atmosphere Networks to Ditech and was chief system architect at Enlighted before Siemens bought it - a company built on data from millions of sensors, which is to say he has been living in the sensor-and-model world for a long time. He did hardware at Cisco before that. The CTO, Jeremi Wojcicki, holds a PhD and runs the data-science and software effort on the model-building approach itself. Jon Gettinger, the CRO, handles go-to-market, having previously scaled Pipe17.

The proof, so far, is in who is willing to put their name on it. In September 2025 ModelCat launched eIQ Model Creator - a tool for NXP Semiconductors, powered by ModelCat, that lets NXP's customers spin up custom models for its i.MX processors and microcontrollers. In January 2026 it did a similar thing with Alif Semiconductor, wiring itself into Alif's Ensemble microcontrollers so developers can go from raw data to a deployed model in under 30 days. Notably, ModelCat says it integrated Alif's chip architecture in under 30 days itself, using AI-driven chip profiling - eating a bit of its own dog food, or cat food, in public.

When two chipmakers decide to hand their customers a third party's model tooling rather than build it themselves, it says something about where the hard problems now live. Silicon vendors are very good at silicon. The messy, iterative, hardware-specific business of producing a good model for that silicon is a different muscle, and ModelCat's wager is that it is easier to partner for it than to grow it in-house.

What can you actually do with it? If you are building an ML-enabled product - a smart camera, a wearable, an industrial sensor, anything in the vast and unglamorous world of the internet of things - ModelCat is meant to take the single most painful and unpredictable part of that project and turn it into something closer to a service call. You bring the data and the target chip. It brings the model. The promised payoff is not just speed but cost, and the ability to ship at all.

The healthy skepticism to hold onto: this is a roughly 13-person company making large claims against enormous incumbents and a field crowded with names like Edge Impulse, Deeplite and Nota AI. "Hundreds of times faster" and "near-zero error rate" are the kind of phrases that demand a field test. But the shape of the bet is legible and, in its way, unusually honest. The company that spent ten years learning how impossibly hard it is to fit intelligence into a tiny, power-starved device turned around and automated the hardest part of doing it. That is a reasonable thing to have learned from a decade of making chips.