Bootstrapped a Generative AI Startup First

I publish this series to discuss the nuances of bootstrapped entrepreneurship. Please subscribe to my Best of Bootstrapping series to never miss an article.

Gridspace Co-Founder Anthony Scodary bootstrapped first, then raised money later to build a $10M+ Generative AI Startup. Gridspace?is a wonderful case study of a speech technology company on the bleeding edge of Machine Learning and Generative AI. You will learn how the founders managed to bootstrap to large paying customers and then raise strategic funding. You will also learn the nuances of how they used various Open Source components and existing ML models to get to a point where they can afford to develop more original technology. You will also learn the importance of solutions versus technology platforms.

Sramana Mitra:?Let’s start at the very beginning of your journey. Where are you from, where were you born, and raised, and what kind of background?

Anthony Scodary:?I was born in Ohio in Cincinnati, although I grew up in St. Louis, Missouri. I went to Stanford for school. My undergrad was in Physics and grad school was in Aerospace Engineering.

Sramana Mitra:?What did you do? When did you graduate from Stanford?

Anthony Scodary:?2008.

Sramana Mitra:?What did you do after that? That’s right at the financial crisis point.

Anthony Scodary:?Yes, but I had fortunately already snagged a job. I went to work at NASA Jet Propulsion Lab (JPL), and it was very similar to what I had worked on in school, which was scientific instruments for planetary space missions.

Sramana Mitra:?That was in Southern California.

Anthony Scodary:?Yes. JPL is part of Caltech, it’s a federally funded research and development center that Caltech runs for NASA.

Sramana Mitra:?How long did you do that?

Anthony Scodary:?I was there for close to six years. While I was there, I worked on the Mars Curiosity rover, Juno mission to Jupiter, and Phoenix lander to the Martian Arctic. I also worked on an unmanned aerial vehicle that flew into the Arctic.

Sramana Mitra:?Fabulous. Then what did you do in 2014?

Anthony Scodary:?A little bit earlier than that, we started Gridspace with my co-founders Evan Macmillan and Nico Benitez, all of whom I know from Stanford. It was focused on machine learning for long-form conversational speech.

Sramana Mitra:?What was the genesis of that? How do you switch from rocket and space stuff to machine learning?

Anthony Scodary:?My background is essentially robotic autonomy, using signal processing for scientific signals. A lot of my work was to get scientific instruments working on planetary space missions. You can only command these spacecrafts once a day or less. Juno was on Jupiter, Curiosity on Mars, which were a long distance away. There’re a lot of missed opportunities because the spacecraft is mostly not doing anything and you’ve limited energy and other consumables. But, ideally, you would have these robotic platforms making scientific decisions about what’s interesting or worth pursuing.

So, I started to learn more machine learning, which around 2010 was emerging as a potential commercially interesting field. AI has gone through peaks and troughs since the 1960s when the term was coined. Notably, there was a lot of excitement about AI in the sixties and the eighties. Machine learning started to emerge as being commercially interesting, partly around the challenge of Netflix recommendation systems. The idea was that you would do a lot of feature engineering, and then build systems that could recommend products or movies. This used classical convex optimization techniques like SVMs [Support Vector Machines].

Sramana Mitra:?The first recommendation engines were collaborative filtering engines for Amazon, right?

Anthony:?Yes, Amazon and Netflix were the pioneers in commercializing that tech, but the problem with those sorts of algorithms is, they don’t work well on raw signals from nature. I was working at JPL on high-dimensional signals where you have maybe tens of thousands of samples or an image. I had learned about some of the work going on with noise reducing auto encoders and deep neural networks, which at the time was very obscure.

Ian Goodfellow, who was at Stanford with me, had invented the generative adversarial network, which was being used for very powerful image generation. I started looking at deep learning as an area that would allow you to operate on high-dimensional signals. Now, it was hard to do that on a spacecraft platform, because it didn’t have the power to run a deep neural network; to be honest, even GPUs at the time barely could.

Even though a lot of these algorithms like radiant descent and convolutional neural networks could be traced back to the nineties or even late eighties, it was mostly hardware-limited. But you were starting to see models doing interesting things that we had never seen in machine learning, in terms of high-dimensional signals.

When Gridspace was founded, we learned quickly that in speech recognition and speech science, there were only a handful of big players. They generally focused on short-form speech like an interactive voice response system where you call into a bank and get this machine to check my balance. Then, you had the voice assistants, a couple years later, like Alexa and Siri.

We were excited about 10-20 minute long conversations, like in contact centers or in businesses, where that information was just uncaptured. Literally every word millions of customers were telling them was completely dark data. They were recording it, but they were doing nothing with it. At the time, there were no speech recognition engines capable of processing conversations like that. Deep learning paved the way for us to do that. Nico, Evan and I became interested in seeing what we could do. We went through Stanford’s StartX program, and they connected us with Stanford Research International (SRI). They have played a really significant role in the history of technology including speech recognition.

Sramana Mitra: That’s Siri.

Anthony Scodary: Yes, Siri gets its name from SRI. They also spun out Nuance. They invested their intellectual property in us. They also gave us a lot of advice on pursuing modern speech tech. At the time, we had used their engines like Decipher and Dynaspeak, which trace their roots back to DARPA projects in the eighties. They were traditional speech recognition engines – Gaussian mixture models, acoustic model and statistical language models. They didn’t use neural networks.

So we had to kind of pave our own way for using neural networks. At the time, the leading open source speech recognition engines like Sphinx out of CMU were not using neural networks at all. 2012 was the first time Microsoft published trying to use a neural acoustic model. It was a very simple neural acoustic model. At that time, it was fairly obscure to use deep neural networks commercially, much less for long-form speech, but we were pretty committed to trying to make that happen. We started building out our first models, and we started getting some of our first customers.

Sramana Mitra:?Were you targeting the call center industry?

Anthony Scodary:?We were targeting basically large businesses that had a lot of voice like financial services. USAA was one of our first customers and investors. Other early investors were Wells Fargo, Santander, and Bloomberg.

Sramana Mitra:?Customers or investors?

Anthony Scodary:?Both. USAA is a big customer and investor, for instance. We’ve processed all of their speech data for four years. They were essentially folks that had very large amount of voice data and particularly cared about the quality of their conversations. This sort of observability was something that I think a lot of people knew they wanted and were very tech leaning. So USAA, for instance, invented check scanning internally, have the patents on it. They’ve always been a very forward looking, customer-focused company. For them, the call center isn’t like a cost center. It was an opportunity and it was untapped data.

A lot of our early customers were very forward looking. We were a lucky machine learning startup that was saying that we could give them access to every word, emotion, entity in their call center calls. We launched that with a product called SIFT, which was our first product.

Our conversation continues here.

Looking For Some Hands-On Advice?

If you’re looking to work with Sramana Mitra?in private,?please check?1-on-1 Consulting. If you are comfortable discussing in public, you can also come to our Free Public Roundtables.