Ali Farhadi is a canary in the coal mine of enterprise AI. His startup, Xnor, is among the pioneers looking for sustainable revenues in neural-networking software for embedded systems.

Today, Xnor launches AI2GO, an offering that consists of hundreds of pre-trained models tailored to run deep learning on a variety of Arm, FPGA, GPU, MIPS, and x86 processors. The 50-person startup, formed in 2017, is already running cash-flow–positive on a small set of early products, but the big challenges are still ahead.

The chief concern is that AI may pan out like IoT — a vast but slow-moving market in which every company needs to carefully think through a business plan that typically requires a custom design. “It’s challenging to get big enterprises moving — they are moving, but slowly,” said Farhadi.

So far, he said, “everyone has their own algorithms, data sets, and constraints. So the challenge is if we can find solutions that scale to many users without much human effort in the loop. On the surface, every project looks custom, but the core technology should be fairly repeatable under the hood.”

Initially, AI2GO targets a handful of what are expected to be the hottest markets in enterprise AI, such as aerial and industrial surveillance and monitoring and home security. Its code runs on cellphones and laptops as well as constrained embedded systems.

Farhadi is well aware that dozens of chip vendors are already offering or designing deep-learning accelerators for the edge. “We offer solutions for a few of them … and we can adapt to new hardware in a fairly automated, fast way, so it’s a great opportunity,” he said.

Xnor’s software lets users specify a deep-learning task such as identifying images of an object. Using sliders, they can select desired levels of latency, memory, and power consumption. Xnor configures and delivers the pre-trained models in minutes.

Ali Farhadi Xnorx

Ali Farhadi (Source: Xnor)

Pre-trained models for popular use cases will be fine for many users, Farhadi believes. The startup is also offering training services for some large customers, but it has not opened that service to developers yet.

So far, Xnor is finding enough work to expand its 50-person company and leave untouched much of the $14 million-plus that it snagged in seed and Series A funding to date. But it’s still early days.

“This is an experiment,” he said. “We think there are embedded developers who want to deploy solutions big and small.”

It’s the first big experiment outside the lab for Farhadi. He studied computer vision and neural networks for nearly 20 years. He still holds positions as an associate computer science and engineering professor at the University of Washington and the Allen Institute for Artificial Intelligence. With Xnor, he finds himself digging an entirely new mineshaft, and the market is watching.