LexSet: Spatial Understanding Powered by Synthetic Data
LexSet is a visual AI solution for object recognition and visual search. I spoke to LexSet Co-Founder and CEO Leslie Oliver Karpas to learn more about what this means, as well as what's currently going on with the company and where it's going next.
How did LexSet come to be?
LexSet doesn’t have the typical startup origin story. The company was born out of the Seattle based research lab, Intellectual Ventures; who’s known for inventing powerful technologies and pairing them with visionary entrepreneurs. In the spring of 2017, I was recruited to IV by now LexSet CBO Azam Khan who was at the time IV’s Director of New Ventures. Az hired me to go through IV’s invention portfolio and come up with interesting startup ideas based on the IP (before Les hired Az to leave IV and join LexSet).
I constructed a vision from a series of patented inventions he found at IV and that IP was directly responsible for inspiring LexSet. My career started in architecture before he moved into tech, and there were fundamental parallels he could see between the technology the patents described and what he knew about interior design.
Beyond architecture, I have deep experience in both 3D technology and the business side of running a startup. I cut my teeth engineering sculptures for the artist Anish Kapoor, and has leveraged his skill with machine vision, robotics, 3D printing, and industrial design in his endeavors since. An evangelist of mass-personalization, I met CTO Francis Bitonti on the 3D Printed Medtech lecture circuit, where he was preaching the integration of machine vision, artificial intelligence, and on-demand fabrication. He started a company called Metamason that custom-printed CPAP masks from silicone.
Francis Bitonti, LexSet’s CTO, is a generative design celebrity and 2015 Wired Innovation Fellow. He’s an award-winning computational designer acclaimed for groundbreaking use of generative design of 3D-printed products, some of which are showcased in the Smithsonian, Centre Pompidou, and other museums. I was a devotee of his design work long before we became friends. Before joining LexSet, Francis built a world-class software consultancy at Studio Bitonti, for clients including the world’s greatest fashion houses, as well as various Fortune 500 companies.
Before Az left IV to become LexSet’s CBO, he was the Deputy CoS of the US Patent Office, a Senior Advisor to the US Commerce Department. He has also served in a variety of political and diplomatic roles including as an advisor to EUSR in Kosovo and a number of roles on the 2008 Obama Presidential Campaign. An attorney and dealmaker, Az is experienced at investment and technology licensing deals, having overseen spinout and startup creation at Intellectual Ventures.
About a year ago, in the spring of 2018, LexSet formally spun out of IV, becoming its own entity.
For those who don't know, what is LexSet?
LexSet is a visual AI solution for object recognition and visual search. Our clients range from furniture retailers to industrial and robotics applications. In short we’re good at recognizing and describing objects in images.
The most common use for our API is to enable retailer’s shopper’s to be able to search with photos, for items that look similar to what’s in the scene.
We’re also being used to automate creating ‘shop-the-look’ experiences. For example, when you’re looking at inspiration photos of beautiful interiors on a furniture company website, we can make all of the items in each scene directly clickable for purchase. Many companies have these experiences, but they have to pay web developers to manually build them, limiting how many they can have, and requiring them to be rebuilt every season; with LexSet- they can have as many as they want, and change them whenever they want for no additional cost.
How does it work?
LexSet’s super-power is building AIs from synthetic data. Let’s unpack what that means: most visual recognition AIs are generated from networks of photographs, which works well as long as there are lots and lots of photos of the content, from different angles and lighting conditions- hence the rise and success of visual search AI in the fashion space. However there are lots of environments and content types where there’s not enough data; that’s where LexSet comes in. We’ve pioneered a technique to create ‘synthetic data’ from 3D models of objects, where we take models of what we’re trying to recognize and render them thousands of times in all possible lighting conditions, backgrounds, and camera angles, giving us near infinite data with which to “teach” our AI. This is why LexSet is moving beyond furniture, where we have clients in robotics, oil & gas, and fabrication all who need object recognition and have incomplete data sets.
How big is the team? Any positions you're looking to hire for in the coming months?
Right now there are 7 of us, who are mostly developers focused on machine learning. We expect to be bringing on additional developers in ML, as well as talent in data science and UX by the end of summer.
What kind of success have you had?
We have a major contract furniture player utilizing our tools for internal use to help them save massive labor hours in the contract bidding process. Our robotics and industrial clients have been very happy with our work to date, and continue to ask us for more. We have a number of pilots with major brands and industrial companies ongoing, and are enthusiastic for those to mature into longer-term relationships as the year continues.
Where does the company go from here?
In addition to creating synthetic data, we are starting to curate our own sets of proprietary data that will further enhance our existing magic. The purpose of which is to get our AI to a state where it can ingest data from journalism without human intervention. Once this is true, we’ll be able to feed the AI content from the entire canon of design and industrial journalism; which is one of our long term goals harvesting historical data for the democratization of knowledge and skill.