Company Profile

COMPANY:CrowdAI

HEADQUARTERS: Mountain View, CA

YEAR FOUNDED: 2016

2016 REVENUE: $35,000 *

EMPLOYEES: 7

More IT Services Companies
  • #4420DSYS
  • #3785TAISTech
  • #3289BIAS
  • #4837Afidence
  • #3084VIA Technical
  • Editor's Note: Inc.'s 12th annual 30 Under 30 list features the young founders taking on some of the world's biggest challenges. Here, meet CrowdAI.

    To do their job, the algorithms that help self-driving cars avoid trees and pedestrians must first know what a tree or a pedestrian looks like. They need a layer of computer vision, and perfecting computer vision requires massive amounts of so-called training data. Basically, someone needs to draw outlines around everything in the computer's field of vision and say: That's a tree, that's a pedestrian, that's a car, and so on.

    That was the kernel for what would become CrowdAI, a deep-learning startup out of Mountain View, California, that today counts Udacity, Planet Labs, and Cruise Automation as clients and Yahoo founder Jerry Yang and SV Angel among its investors. Although it only started generating revenue last year--the revenue figure reported in its 30 Under 30 application reflected sales closed while still going through Y Combinator--CrowdAI is well-positioned to capture a large piece of the fast-growing market for labeled training data.

    Identifying an advantage

    Devaki Raj was working at Google, as a data scientist in the energy division, when she came to appreciate how crucial training data would be to unleashing the power of artificial intelligence in the world. In conversations with one of her colleagues, Pablo Garcia, and a former colleague of his, Nic Borensztein, she puzzled over what kind of company they could start around that need. "We were all really motivated to get involved in this new wave of AI and think about how we can solve this computer vision problem," says Borensztein.

    A number of companies were already tackling the training-data-shortage problem by using huge networks of people to annotate images. Raj, Borensztein and Garcia agreed most of that effort could in fact be eliminated with yet another layer of AI, called deep learning. "If you can get enough examples of humans making a decision, the AI acts as a black box to reproduce that decision making," Borensztein explains. They conceived a platform wherein AI does the easy work and humans the hard parts, with the burden shifting more and more to automation over time as the software gets smarter.

    Gaining clarity

    When the three entered Y Combinator in 2016, they knew the kind of technology they wanted to build but hadn't decided what specific business applications made the most sense to start with. They toyed with the idea of focusing on speech-to-text transcription for doctors who dictate their notes, but their advisor, Y Combinator partner and COO Qasar Younis, suggested any product catering to hospitals would be slow to get its first sales. "YC beats it over your head: Make revenue, make revenue, make revenue," says Raj. "It helped structure our thinking in a way that we could build a financially sustainable company." They turned their thoughts to the burgeoning autonomous-vehicle market, where they found one of their first customers, Cruise Automation, the self-driving tech startup acquired by GM in 2016 for more than $1 billion.

    Around the same time, they found a second specialty in using human-corrected deep learning to annotate satellite images. If, say, a hedge fund wants to get a sense of whether foot traffic to a big-box retail chain is up or down, it can send overhead imagery of parking lots to CrowdAI to get a handle on how busy its stores are from month to month. "There's so much data to go through, it's not possible for humans to do it alone," says Garcia.

    Younis says he was struck by the "quiet confidence" of Raj and her co-founders. They're far from the only startup using machine learning to annotate the world's visual data, but great ideas are seldom unique, he says. What matters are structural advantages like great technology and superior engineering talent, which CrowdAI has, he says. "They can actually do things other people can just not do."

    Going the distance

    Even so, the pragmatic Raj is hardly dismissive of her company's challenges. "A lot of companies have the resources to go into some of the work we're doing," she acknowledges. But the big companies, like Google and Uber, aren't going to share their proprietary technologies with each other or with upstarts trying to overtake them. Nor, she insists, are they going to be as good at it as a startup that focuses maniacally on doing one thing better than anyone else. CrowdAI's algorithms already are able to annotate 70 percent or more of the images fed to them without human assistance. "We're very good at deep learning from satellite imagery and very good at road imagery," Raj says. "The marriage of the two is where we think we can find a very large place."

    Currently, a self-driving car has to use its radar and laser sensors to grope its way over every object in the world around it--cars, pedestrians, trees, lampposts, etc. That eats up a ton of computing power. Why not use satellite imagery to pre-map the location of stationary objects, freeing up memory for more urgent calculations?

    Ultimately, Raj says, CrowdAI's vision is much bigger than cars or satellites. One thing she's particularly excited about: the idea of using machine learning and computer vision to help understand environmental change and cope with natural disasters. "There are so many questions to be answered with satellite imagery, and some of them could certainly benefit humanity," she says. "We would like to be a flexible deep-learning platform for external objects."

    CrowdAI - 30U30
    Published on: Jun 5, 2017