Launching Astra: How Deep Learning helped us launch our Financial Intelligence startup
Every startup has a unique origin story. Certainly there are some common themes — a genius founder, recognizing a great trend, or just grinding tirelessly out of a garage. But what if you have an idea for a company without all the key ingredients on-hand? You are passionate about the problem, and you have ambitions to achieve significant long-term impact through technology. But how do you get started? And are there examples out there to guide you?
This is the origin story of Astra, of the struggles all of us have with the current banking system, of the financial intelligence we think the world needs desperately, and of how my co-founder and I charted a course towards our first product and foundational technology — without any prior domain expertise. Unlocking the benefits of Deep Learning with a Nanodegree program from Udacity gave us the skills we needed to apply Artificial Intelligence to personal finance, and enabled us to get our startup off the ground.
This is the first in a series of stories we will be writing about our experience starting Astra. Follow me to receive a notification about the next chapter.
Two years ago when I was living in New York City, my friend Sam came through town and was looking for a place to crash. We met at my apartment, took in the night skyline, and toasted to the opportunity to catch up. I had just spent the past few days deep in spreadsheets modeling the intricacies of my company’s finances, and he was in the midst of modeling the impact of whether he should take a new job in a new city — with all the different fixed costs, variable costs, cost of living, and other options. We ended up having an impassioned conversation deep into the night about the shortfalls of the financial services and tools available to us. We both had steady jobs, and might actually be making progress towards paying off our debt. But it was hard, even with our custom tools, to be sure we were on the right track.
Identifying the Problem
What we ultimately realized, was that while we had each created improvised tools that put more information at our fingertips, my financial model — filled with hacks to my Mint account and complex spreadsheet functions — wouldn’t translate to Sam’s, or vice versa. We also couldn’t take our projections to the bank and say, “See? You actually do want to lend me money so that I can refinance my student loans.” We certainly couldn’t use them to improve our credit scores either. We were empowered with real data that was helping us make better decisions, but our models were only truly meaningful to us — no one else cared.
What we did realize, though, is that we were both passionate about making these kinds of tools. Independently, we had both been striving to cobble together a full picture of our finances. It was never explicitly stated in that conversation, and even though we both knew what we had would probably just gather dust, we also knew that smart models might become special. Sam departed the next day, but the conversation didn’t end, and neither did the challenges of managing our finances.
Over the following few months, we accumulated stories from our friends. They’d tell me about their finance hacks, or tell Sam about the ad hoc tools they wired up, or tell us both about how unhelpful their banks were. A buddy from college shared that his tech startup had low overhead, $50,000 cash in the bank, a $500,000 contract with a Fortune 100 company, and he still couldn’t secure a small line of credit to grow his team. Despite the uptick in the economy after the worst of the financial crisis, it didn’t seem like anyone I knew was gaining any ground.
The average savings rate for people under 35 is negative 2%. Home-ownership for our generation is at an all-time low of 28%. And student loan debt has surpassed $1.2 trillion — thanks in small part to my own contribution. Everywhere we looked, we saw a divide. On one side: where everyone wanted to be with their finances. On the other: where everyone actually was. Surely there was a way to bridge this gap? Our one-off models weren’t the answer. So what was?
Spend zero time on what you could have done, and devote all of your time on what you might do.
Searching for a solution
What we needed was a more powerful tool, with much more range and adaptability. If we were reading the stars correctly, we needed to be able to see if finance really could be smart, and more personal. We didn’t know what that tool was yet, but we knew we had to build it and make it work for everyone.
We needed a technological solution that could incorporate the unique aspects of each person’s situation and spending behavior, but generalize well for the broader challenge of making finance smart. I knew from my prior technical research and development experience that even the more advanced approaches within the algorithmic paradigm — some even drawn from the complexity of nature and evolution — couldn’t accommodate such a range of inputs. We needed systemic smarts, that we could use, test, and deploy.
I had been casually tracking the developments in the field of Artificial Intelligence, and I’d read up on TensorFlow. It seemed like financial intelligence could become a thing, but we weren’t sure — and even if it could work, we didn’t have way to know for sure, much less to make it real. We felt like the stars were aligned, we believed we could positively impact people’s finances, and we wanted to test our hypothesis about smart finance, but how?
As soon as I discovered Udacity, I started binging on every free and supporting course they offered. After a couple weeks of this, I made the decision to enroll in the Machine Learning Nanodegree program. As I worked through the curriculum of supervised, unsupervised, and reinforcement learning, it was clear that all these systems were entry points for creating the type of intelligence we need to improve our financial health. Beyond the technical content, every project submission and review by Udacity’s program mentors was valuable applied experience. Even though trying to predict housing prices in Boston or training a self-driving car to make its way through a maze weren’t directly applicable to helping me design smarter finance tools, those projects helped me assess different categories of machine learning.
For my final project, I developed a Deep Learning Stock Price Predictor to see what type of model might best predict the value of a volatile asset over time. The results, and the Long Short-Term Memory Neural Networks that comprised them, became a key ingredient of the financial intelligence we’ve developed since, including the ability to predict the balances of your accounts in the future.
Starting the Nanodegree program was meant to fill a technological gap in our equation. We bridged that, and then some. Udacity gave us the skills to build a more powerful tool — a telescope. An instrument that lets us see financial stars more clearly, and interpret them more confidently.
Sam and I founded Astra as a financial intelligence company with a mission to make your finances more tangible, empowering, and personal. We believe that developing smart technology, powered by deep learning, can help everyone improve their financial health. A few months later, with a beta in hand, we’re ready to share what we’ve built.
Check out our beta and let us know what you think. Follow me to receive a notification about the next chapter in this series.