Who are you?
My background is in synthetic biology. I spent around five years working on big data associated with genomics. I recently started a company called Pavlov and we're trying to make machine intelligence more accessible to everyone.
How did you come about with the idea of this company?
A lot of the problems that me and my co-founder, Alex Kern, set out to solve were problems that we faced ourselves and problems that we saw both researchers and industry engineers facing themselves. One of the key problems was a lot of machine learning technologies were research efforts. There didn't seem to be an enterprise-supported landscape.
We started working a lot of projects where we would try to use all these technologies out of the box and since they were research efforts, a lot of times these products weren't documented well. They didn't really work out of the box and it required a lot of handling to get these things operating and running. When I worked at Harvard Med School, I quickly realized there were very few people in our lab, full of MIT PhDs and Harvard MDs, who could use big data technologies.
The technologies that were developed for big data, namely Hadoop and Spark, etc., were built out of necessity at companies like Yahoo and weren't really engineered for general purpose use. This is kind of the problem that we saw that exists in the world and we also quickly began to realize there are a lot of advances happening in machine intelligence, specifically the computer vision space.
We're seeing orders of magnitude changes in performance and accuracy in the kind of technologies that we have available to us over 6 months to a year's span so this entire landscape is evolving dramatically.
At what stage is your company? Are you launched? Do you have customers? Do you have revenue?
We started the company about 8 months ago and the company itself has launched. We went through Y Combinator’s first fellowship batch. We pushed off participating until the fellowship came up because that was a unique experience for us. It provided us with a small, equity-free grant and gave us a lot more ownership over the focus and direction of the company.
Since then, we have the enterprise customers and direction. We have about 10 letters of intent. We have a core product offering and we have unique use cases that we can scale out as the business grows.
How do you sell Pavlov to these enterprise customers?
This is an interesting question and a really good one. I feel I've learned a lot about how to sell a machine intelligence product in an enterprise market. It's a little different than just selling a data analytics product mainly because the solution is a more complex solution. You're not providing just one part of the puzzle. You're providing a whole system.
We have an elastic, on-demand work force of humans that we're able to scale, along with the computational resources to train and seed some of the model training.
We provide a lot of these solutions to enterprise and what's selling this to a company looks like is really going to an executive or a high-level engineering manager and present to them a very unique value proposition. Rather than hiring a data scientist or a data science team, would you rather try out our solution and enable existing engineers to perform the tasks that are often presented as unique. We found that engineering managers especially are desperate to maintain a head count for core engineers.
Solving this huge problem, providing the tooling, providing the support, and a managed pipeline for them to perform machine learning at scale really allows them to not even have it on pause but double down on the engineering headcount and resources in other areas of their company. One's far more critical to their value proposition and core product offering.
It's interesting as well to look at what the sales cycles look like. We are engaged in a few different spaces. We originally started this company with the goal of making sense of satellite imagery. We have a few deals that have given us terabytes and satellite imagery and we're applying learning to this data.
The sales cycles presents this data to customers that find it valuable. It's somewhat longer than we expected. In the meantime, we've actually approached a shorter-term solution that a lot of those who delivered apply the same computer vision technology to do something as simple as copyright detection of a type of fraud that is often overlooked but really relevant among customers. Right now, we've launched something called Pavlov's Bot. There was a research paper two months ago and it was the basis on which we developed this algorithm. We're selling to big companies that have huge user-submitted images and creative art work. You can think of Etsy and CaféPresses of the world. We're going to companies and these companies often have users submitting copycats, impinging on other users' creative works or actually impinging on the creative trademarks and copyrights of big companies like Coca-Cola and Heineken.
Our technology is able to capture 80% of these brands and logos and mitigate legal action for some of these brands. It also remove a process that's highly human-driven. These companies have teams of 18-25 people that are working day-to-day to review every single user submission.
Currently there is no automation process. We are able to go to these companies and give them a multi-step solution that allows them to slowly transition from a fully-human solution, manual process, to a somewhat high-functioning, automated machine-intelligence solution triggered by a few humans that are playing almost like an account management and escalation triage role.
It's interesting selling this product. I think it varies by market as to how to sell the product. One last key insight we found is depending on the area and ML or AI that you're focused on tackling, it's often a great idea to evaluate which companies are trying to hire very unique ML or AI talent. Approaching those companies seems to fast-track our sales cycle tremendously. These companies are not only looking for the talent but are actually trying to invest in the technology stack and are more than willing to provide the tooling necessary for the existing team or the new team to start becoming invested in this space.
It looks like you're not just providing tools, you're also providing applications or application frameworks. Can you say something about that? How much consulting do you do?
We actually started off with a core vision of never doing consulting and not providing application-level solutions. We had a strong belief that we were merely providing the tooling necessary to allow this business to scale from an infrastructure perspective and not one that was bogged down in providing big enterprise contracting solutions.
We realized after a month that was impractical. On one side, we didn't want to be a big consulting company, a Palantir or a Databricks of the world. On the other side, we realized it's really hard to generalize ML and machine intelligence. We also realized that a lot of the learning that we've made in terms of core product and infrastructure have come directly from our customers and without having customers to deploy in production, it's very difficult to get the feedback that someone is trying out your system, playing around in the system, is able to provide. It's just not as important to them.
Therefore, we kind of went about coming up with an approach that allows us to build small solutions as a stepping stone. We decided to build Pavlov's Bot primarily because there was a very clear business usage to one of our customers, one that was willing to pay a tremendous amount of money for us to invest only a few weeks to build them a custom solution.
In doing so, we actually created a system and a product that's repeatable. We own the model and we're able to provide that model to many, many businesses and this is something where we were able to build a unique use case and application and monetize it at scale. We still provide our full-stack core product, Pavlov. It's provided to most companies as a push to deploy full-stack machine intelligence solutions and let them make what they want.
Having this two-tiered approach to business development has allowed us to fast-track our product development cycle, while allowing us to continue development towards a long-term automated solution that we want to be able to deploy in big companies with very little implementation effort.
We addressed the need to scale the business from two people. With a team of two people, taking on new contracts has been a challenge, especially when integration and implementation is often quite a lengthy process. By building Pavlov's Bot we were able to hire consultants and contractors. These are akin to core deployment engineers that have a very clear goal. Take core products and implement it with a given client. We were able to tie this into the pricing structure of our enterprise sales contracts and it allows us to quickly build a strong enterprise solution while still trying to drive towards a more generalized offer.
How did you acquire the skills needed to do AI and run a startup? What is your educational background and professional background? What interesting stories do you have?
I grew up in South Bay, moved to the East Coast when I was in high school, spent around five years working at Harvard Med School and over this time, I worked in the Church lab trying to tackle really hard synthetic biology challenges. These challenges presented a lot of data, data that many scientists in the lab had no idea how to make sense of and neither did they have the tooling necessary to process and distribute the computational resources necessary to solve them.
That's where I got my first exposure to computer science and big data technologies. Fast forward a couple of years, I was in college in Albany in an accelerated medicine program (I participated but did not complete it) and I realized that the problems we were trying to solve were ones that were much more focused on data, rather than being a primary care physician. Took a break from college, moved out here to join a startup.
After that, I met my co-founder, who previously built infrastructure both at NASA's JPL and Apple. We kind of built their entire product infrastructure. Working with them, we both decided that machine learning and a lot of the processes that existed needed some degree of abstraction, an abstraction layer.
We created a core language. That's how it all started. This allows you to manage the entire infrastructure from deployment to data science, all from a simple query interface. It helps them answer hard questions about their business as quickly as possible.
Can you summarize the state of AI hardware. If I want to get 100 GPU's real fast and real cheap, what do I do?
That's a great question. The scene of ML hardware is one that is quickly changing but without very easy solutions to the problem. There is no Amazon of GPU research. What Amazon did by democratizing CPU resource, that's not been done for GPU's.
We've been looking and looking and looking from IBM software to any vendor who wants to sell a co-located box and what we're realized is some of the biggest companies in this space are buying boxes, building boxes and co-locating them. This is why at Pavlov, we actually have an open-source design for a 5U unit. We're not really selling these boxes other than co-locating them currently at a few data centers and offering them as a hosted service to our customers. We're talking to Fortune 100 companies that struggle to understand the concept of building gaming rigs to help them do some data science.
These companies aren't even in the cloud. We bring to them a reassuring solution when we offer them a 5U unit with as many GPU's as possible with software we've written that optimizes the jobs and resources. They're actually able to connect desktop monitor, keyboard, and mouse and jump right into an interface and start playing around with Pavlov.
For these companies that don't have our support and don't really want to build a bunch of gaming rigs, this is the only solution that we could come up with to provide them what they want. As to the state of disarray, the fact that I have to build hardware when that's not even part of my business model and not something I'm really interested in doing, has made me realize that these expensive processors act as a barrier to entry to many people interested in leveraging these technologies.
Why do you think there's so much excitement about AI now?
Honestly, I have asked myself this question a lot. The reality is any company of a tremendous value has a lot of data and making sense of this data used to be an afterthought. It's now critical to product experience and business strategy to make sense of this data. That's the first reason why there's going to be a huge rush of ML applications and solutions that are going to hit the market.
Another key insight that I found, making sense of all the archive papers that have been published, and doing some degree of clustering and analysis, a lot of the recent advances in ML have not actually been in networks, the area where you have virtual bots and NLP being developed. A lot of the recent advances and the highest acceleration of change exists in the space where the technologies that are used to make sense of the images at scale. We now have orders of magnitude better algorithms and technologies and processes that allow us to basically make sense of these huge data sets and do change detection and identify patterns that are previously unfathomable. That's why I think AI is making a huge coming.
It's that the tooling necessary to go from zero to doing some ML is becoming more and more available and it's companies like myself and others that are offering ML as a service. They're just little steps towards the direction where the entire industry is moving.