Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. Benjamin Franklin once said, nothing is certain except for death and taxes.
Actually, it may have been Edward Ward or even Christopher Bullock who first said those famous words, but no matter, they're true.
We all must face death and taxes eventually.
Today, we're going to talk about the scarier of those two topics.
That's right, we're talking taxes. Ashok Srivatsava is Senior Vice President and Chief Data Officer at Intuit, where he's focused on large-scale machine learning and AI.
He also serves as adjunct professor in the Department of Electrical Engineering at Stanford University.
Ashok has previously held positions at Verizon and NASA, amongst other places.
And so I'm not going to let him leave the podcast until we talk at least a little bit about NASA.
But first, we're going to talk Intuit and taxes.
Ashok, thanks so much for joining the AI podcast and welcome.
No, it's so great to be here. Thank you for inviting me.
So full disclosure before we start, I am an Intuit customer.
But I pay for my subscription to QuickBooks Self-Employed, I think it's called.
I'm not receiving any free tax advice. for having you on the podcast, although we can talk later if something comes to mind.
So let's start with the present. You joined Intuit last October 2017, is that about right?
Yeah, it started early October 2017. So it's been a little bit less than a year, and I've read a lot about Intuit using machine learning and AI to help customers.
So can you talk a little bit about that and what your role is as Chief Data Officer?
Yeah. So my role as Chief Data Officer is to help define the strategy for AI and machine learning and data for the company in order to serve customers as best as we can.
And part of that is building a team. Part of that is building technology.
And part of that is also understanding the customer needs and the customer demands.
And so my team works in all of these areas, It's an extremely exciting area because as you know, AI and machine learning is just taking off right now in so many industries.
But what better industry out there than finance to see how these technologies can be used to help really regular people who are leading their lives do better in a financial way.
I'm one of those regular people. And like I said, I use the self-employed product, which helps me track expenses and income coming in and out and that sort of thing.
And I started using it mainly just for the ease of use.
There's a smartphone app. It connects to my... my business bank account and my credit card.
And so I kind of see all the transactions come up and I can mark them as business, not business, that kind of stuff.
But my understanding is that machine learning and AI adds another layer to make things even easier.
So can you speak a little bit about that?
Yeah. One of the things that happens when a customer is going through their daily lives, they have to understand, for instance, things like the categorization of the transactions that they're doing.
So you go to a restaurant today and is that a business expense?
Is it not? You might go to buy some supplies at Office Depot, and you want to know, should I deduct that or not?
All of these kinds of decisions. could be done by humans, but wouldn't it be way better if it was done by artificial intelligence?
And so that's just one example where we're using AI techniques to make people's lives easier so that In the background, the machine can automatically go through all of that data with your permission and then categorize it for you and just give you recommendations.
Another great place is in helping people get credit.
For instance, right now a lot of small businesses have difficulty in getting credit.
And we're building AI techniques and machine learning techniques that can help people get credit.
And as they get credit in their small businesses, helps them expand and helps them grow.
And it really helps build the foundation for society, if you think about it.
A lot of society is built on small businesses and we're trying to help those businesses do better.
Now, is Intuit focused specifically on self-employed individuals and small businesses?
So small business and self-employed is certainly one area, and then consumer is another area.
So those people that use TurboTax or any of our other tax products are falling into the consumer area and they would enjoy similar kinds of benefits.
So helping people fill out their taxes in a more timely way getting better confidence on how their taxes are being done getting questions answered quickly.
And then also making sure that they get the best deductions that they can.
Those are some ways that we're helping people really do better in their financial lives.
And this is the way we think of powering prosperity around the world.
I've always been terrible at taxes. There probably is some sort of a stereotype around... somebody who's worked as a writer and a sort of, quote, creative type and being being bad at being organized and paying the price, at least psychically, if not financially come tax time.
But I've sort of at least evolved a little bit to where, like I said, I use the tool, but I still have somebody who I pay to prepare my tax returns for me and with me at the end of the year.
Are we at the point, and if not, how far away might we be from an AI-driven system being able to suck up all of my financial data from the year and spit out my tax return.
And all I have to do is sign an e-file. Well, we're making our path to that end result.
We're really taking the time to build the technologies that can help people do their financial advising and to get their taxes done as quickly as possible without having a lot of manual work.
But some of the things that always will need to be done are getting some context about your life, the kind of work that you're doing and so forth, if anything has changed, that might not be directly related to what data we have available.
But other than that, yeah, we're moving along that path really well.
We're building systems that can literally ingest data the IRS, this might be tens of thousands of pages of documents, automatically ingest that information and then build ontologies, build other kinds of knowledge graphs behind it so that the machine itself has an understanding of taxes and how taxes should be filled out.
It's a pretty amazing capability that we're building.
And so that kind of leads nicely to what I was planning to ask you next, which is about the tax code itself. and how that relates to the technological challenges that you're facing.
You touched upon the machine learning systems, learning to categorize different types of expenses, and is it even a business expense or something personal that can't be expensed.
I don't know if the right question is, how much is this and how much of that is that?
Or maybe just can you speak to some of the challenges you're facing on the side of the consumer behavior and learning to categorize different things, but then also on the side of the tax code.
The tax code is something in America that untold comedians have made at least parts of their living talking about how complicated the tax code is, let alone the number of CPAs who make their living. because people like me just can't make sense of it, and it changes all the time.
And so how big of that is a challenge, and what are you trying to train your machine systems to learn? and understand?
Yeah, what a great question. So there's no doubt that the taxes in the US change.
And as they change, they're extremely complicated. and it's very difficult to keep up with it.
And so we're building techniques that can actually go through The written code, these thousands and thousands of pages of code, all of the forms, I think there are several hundred forms that go into this system.
And it learns the relationships between the instructions and the forms themselves and what data should show up in the forms. and what data should not show up, Data should be combined together to give you the best end-year tax result.
So it's a system that has both machine learning in it. which means the ability to learn from data, but it also has knowledge engineering.
In other words, building in rules and rule-based systems, expert systems that can take that information and codify it in a way that machines can understand.
Taking these two things together, it's a daunting task, no doubt.
But as the tax code changes, as it just recently did, The systems can keep up and evolve with it so that something new happens and it goes in through the ingestion pipeline.
People have to look at it, but in the end, that code goes into our systems and it goes into production.
Is your ingestion pipeline, as you called it, Is that an automated machine at this point that's using natural language processing to parse the actual English written tax code?
Natural language processing, natural language understanding, and machine learning.
These are the three major pillars on which that technology is built.
The natural language processing and understanding means that We're building systems that can actually read, quote unquote, the English language that's written in these tax forms and in the tax documents and build. a knowledge structure underneath.
That's a very difficult task. And then the machine learning component is it's really novel because what it does is it looks at the data that comes through And it matches it against the information that's encoded in the ontology.
And it says, does this make sense or not?
And it actually gives us the ability to explain why certain tax decisions might be made or not made.
And so It's a very, very foundational element of our entire technology stack.
It is an ingestion pipeline. It is, I would say, semi-automated.
Most of it is being done in an automated fashion, but there are always tax experts and other experts that are looking at it just to make sure that everything is being done really, really well.
In previous episodes of the podcast, I've talked about natural language processing and understanding in the context of AI with... a researcher who was working on a sarcasm detection engine and more recently with somebody who was who had built a neural network to auto-generate memes.
And so as I'm thinking about the challenge of parsing the tax code, I'm trying in my head to sort of juggle like, Well, which is harder to comprehend, sarcasm or the US tax code?
I'm glad there are people like you who are grappling with these things so that I don't have to, frankly.
Our guest today is Ashok Srivatsava. He is the Senior Vice President and Chief Data Officer at Intuit. where he is building and leading teams and setting the vision and direction for large-scale machine learning and AI across the enterprise, helping folks, individuals, small businesses with their finances and with taxes.
And as I mentioned at the top, you previously, I believe you came to Intuit from Verizon.
Is that correct? Yeah, that's right. I was chief data scientist at Verizon.
So you have some experience, obviously, working on machine learning at what we call large scale.
What exactly does that mean? And what, if any, can you speak to... things that are unique to when you get to sort of large scale applications of machine learning, things you've learned along the way, maybe things that surprised you once you got to working at such big scales.
Yeah, absolutely. So I love large scale machine learning problems.
And the reason is it gives you the opportunity to understand huge systems like how people think about their taxes, how people do their finances, how people use the internet, these major, major Areas that people spend a lot of gives you the ability to understand that at a very deep and refined level.
And it's one of the only techniques that I'm aware of that can give you that ability.
So let me give you a real example. Let's say that we have data that's coming in from people who are self-employed.
So these are people who are writers, who might be educators, they might be working for Uber or other companies like that, and they're self-employed, and yet, They live across the country in different areas.
And yet what you find is that they have some similar needs and similar characteristics And there's a unique opportunity to serve those people in a unique way, in a personalized way.
And so large scale machine learning lets us take all of this data and use it to help power the prosperity of one individual.
And so what's exciting about that is that these models take in sometimes thousands, sometimes millions, sometimes tens of millions of variables. reason across those variables in order to give that self-employed individual the best recommendation possible.
So it's exciting for a couple of reasons.
From the technology reason, it's extremely exciting because it gives us the opportunity to use large systems, lots of data, and build powerful algorithms.
But I'd say even more importantly from the consumer aspect, gives us the opportunity to serve that customer in the time and in the context that matters to them most.
So for me, It's incredibly exciting to be able to draw the line from both of these areas and help people do better in their financial lives, power prosperity across the world.
All for that. Let's change gears a little bit and talk about your background.
I know you did your graduate work at the University of Colorado in Boulder, Colorado.
Are you from that neck of the woods or where did you grow up?
Yeah, I was born and raised in Fort Collins, Colorado.
Okay. I am proud to call myself a Colorado native.
So there are a lot of people who have transplanted into the state, but I was actually born there.
I went to CU Boulder for my undergraduate as well as my graduate education.
It was a phenomenal experience. What a beautiful state.
And then you made your way through several different experiences, one of which was a tenure at NASA.
Yeah, absolutely. I was at NASA for a little bit right after graduate school, and then I went back to NASA for I think about 12 years in total.
And it was a phenomenal experience. I absolutely loved the agency and the mission that it had and the people that I worked with.
And so where were you and what were you doing?
So I was at NASA Ames Research Center, which is here in the San Francisco Bay Area.
That's where I was stationed. But I also had a role within NASA headquarters where I would run a large research and development program that was focused on advanced technologies to help make aviation safety more prevalent across the world.
And so it was a huge undertaking and such an honor to be able to work there and to be able to represent the public's interest in aviation safety.
Now, was this big data machine learning focused as well or not yet?
Yeah, it was actually. The research and development program that I ran had many components to it.
One was focused on diagnostics and prognostics. which is essentially trying to understand complex systems and how they might have problems with them. and also how long they might operate before they have failures, human factors, but then there was also another area which was focused on machine learning and what we called data mining and using these kinds of technologies to improve aviation safety.
And there was also a component of work focused on what we called verification and validation of flight-critical systems.
So this is actually... building technologies to make sure that software is stable and secure and safe when it's in operation.
So it was a massive program. I also led the machine learning team here in the Bay Area, which was a tremendous amount of fun.
There we branched out of aviation into earth sciences, space sciences, astrophysics, and we got to do a lot of different types of things in the scientific domains.
So life on Mars, yes or no? I believe that we're going to find that out soon.
Well said. So you've been working with kind of on the advanced edges of data and using large swaths of data for pretty sophisticated purposes at the highest levels in the aviation and aerospace industry. in the telecom industry at Verizon, now in the finance space with Intuit, when you get to these large scales of data and these levels of sophistication,
Are there more similarities than differences, even though you're in fairly different industries?
There are a lot of similarities. So the similarities of the complexity of the data, the way the data comes in, the timeliness of the fact that it's often asynchronous, all of these things I've seen across domains, all the way from advertising to astrophysics. to earth sciences, you know, you name the extremes, I've seen it in all of those domains.
And so that's what makes machine learning so powerful is because it can take advantage of that commonality.
But there are pretty significant differences as well.
And some of those differences make each of these fields extremely exciting and compelling to work in.
So for instance, in finance, It's about some of the most personal decisions that people make throughout their lifetimes.
And so the attention and care that's put into building the algorithms and building the models and understanding the data is really unprecedented in this domain.
In other domains, for instance, if you're working in the earth sciences, it's really not about understanding an individual person's preference or their interests or anything like that.
There it's about understanding large-scale processes that are happening on a planetary basis.
The issues there are different, but the care and the understanding of the data and understanding how the data is coming in and what it really represents is critically important.
So these differences provide a lot of intellectual stimulation that makes it extremely exciting.
And then the commonalities make it so that machine learning and AI can be applied in so many domains and advertising astrophysics all the way to Other extremes like finance or telecommunications, cybersecurity, all of these areas, you'll find that these technologies play an important role.
And whether looking back, even just at the little bit less than a year you've been in it to it, Or do you know over the 20 plus years that you've been working generally in this domain, What, if anything, along the way has surprised you or maybe stands out as kind of a breakthrough watershed moment in the work that you personally have done?
I've been so excited to see how some of the techniques that we've worked on, like anomaly detection, and prediction problems are ubiquitous across different industries, across different customer bases, and frankly, across different sciences.
So anomaly detection, the ability to determine whether something that you're observing now is the same as what you've seen before is different than what you've seen before.
This is like a very foundational problem that you can imagine has a role in understanding the Earth system and climate. you wonder, gosh, it's really hot these days.
Was it really this hot before? Is it really different than what we've observed before or not?
That's an important question. In finances, is our current finance situation, is my cash flow today similar to what I've seen before or is it different?
That's a common thread and understanding anomalies not just based on one variable, but maybe based on 10 variables or 100 variables or 1,000 variables or more, that's an important problem, something that I've worked on with my colleagues for a long time. that I'm super interested in.
Another problem that I'm extremely interested in and one that really compelled me almost to come into the field of machine learning was the science of making predictions.
So we always want to know what's going to happen in the future.
And in fact, if you go back a few hundred years to the time of Isaac Newton and Leibniz, what you'll find is that these scientists and others We're always trying to build models that can help us look into the future.
And some of the work that people have done in machine learning gives us the ability to make predictions and make those predictions with confidence.
And that's extremely exciting, something I've worked on for a long time. but countless other people have worked on as well.
And every day I wake up and I think about new ways to build predictive models, new ways to build anomaly detection models, new ways to understand texts and written language that just makes this such an exciting area to work in.
So to tug on a couple of those threads and bring them back to what you're doing at Intuit, and then of course, bring them back to myself, because that's what I do.
I'm thinking about the self-employed app.
And one of the things that I like most about it is that it gives me a bead on my quarterly tax payments.
As somebody who's been self-employed and freelanced for a long time, You know, there's always this sort of game of, oh, I know I should pay quarterly taxes, but also I know that if I don't, I'll get a fee at the end of the year.
But, you know... not going to get taken away to jail, you know, because I missed my quarterly tax payment.
And so is it better to make the payment or keep the cash for cashflow and make it later, et cetera, et cetera.
And so one of the things I like about the app is that it crunches the numbers and lets me know.
Are there... predictive things that are being worked on or could be applied to the individual or the small business you know, looking at their financial situation.
Are there things that Intuit offers that use AI in this capacity or perhaps things that are coming down the pike?
Absolutely. So let's think about a small business that might be building its business up, hiring a few employees, buying from suppliers and so forth.
And they have some money coming in on a regular basis from their customer base.
And they also have some debt and they also have some expenses.
So this pretty much captures most small businesses within the US and abroad.
Now, if you think about it, one of the key problems that they face is understanding their cash flow.
Am I going to have enough money at the end of the week or at the end of the month to make my payments to my creditors and to my employees?
And so there's a prediction problem in there.
And sometimes business people do that just based on a spreadsheet or just based on some quick calculations they do.
But it's hard to think about all of the factors that actually drive that particular financial decision.
This is where machine learning and predictive models can play an important role because they can look across not just one or two or five variables, but they could look across 50 or 100 variables and say, you know what, you should plan to bring some money out of savings in order to make your next payment to your creditors or the next payment to your suppliers.
Or you're doing really well. Take some money right now and put it in the savings so that you can use that for a rainy day later on.
Those kinds of financial decisions. are things that can be made through the combination of cash flow prediction, which is something that we're working on, to deploy and also it can be used with other techniques like advice and recommendation systems that we could be giving at the moment.
Because as you know, sometimes a recommendation that you get you can act on right away, and other times you might get a recommendation that you can't act on.
And we want to know exactly when is the best time to make that recommendation so that you're most likely to get the best financial advice and make the best financial decision.
So that's one example, but there are many other examples that I could go over.
So for folks, whether individuals, small business owners, or anybody else who might be listening who wants to find out more about what's available now from Intuit and things that you're working on.
Are there places online they can look? Yeah, so we have... a blog that's published, and we can give you a link for that, where we talk about some of the technology innovations and product innovations that we're doing on behalf of our customers.
We also have a good deal of material that's coming out that talks about specific use cases through other channels, podcasts, and other things that we've done, which we can share with you as well.
Excellent. You know, I'm so extremely optimistic about the use of AI and machine learning to help power prosperity for people around the world.
If you think about what we're facing today, a lot of the societal issues that we're facing have an economic and financial underpinning to them.
And we really feel that taking the combination of people, of data and of AI and machine learning and bringing these things together can really help address some of the most basic things that people are facing from a financial perspective.
We feel that if we can take that and help people make better financial decisions in their own personal lives, it'll help everyone around us.
And that's essentially the mission that we're on.
A follow-up on that. When you talk about reaching everybody in the world, or many people around the world, I'm thinking of folks who may not have access to, or I don't know, I guess I'm wondering, you know, as I mentioned, I have... you know, an app on my phone and I pay what to me is a very reasonable fee to use it.
And it connects automatically to, my bank account, my credit card account, and some other accounts I'm sure.
How much of the world has that sort of ability right now to leverage the automated systems to connect the account so that the data can be pulled in and examined and the recommendations can be surfaced back to the individual to help them.
And how much of the challenge... for you in that mission is connecting more of the world's individuals.
The number of mobile phones and cellular devices that are out around the world is extraordinary.
And these are not just the old type of phones.
Smartphones are ubiquitous and that gives everyone an even playing field, which is very, very exciting.
And so Intuit's mission is to build products that can help people around the world do better from a financial perspective, empower prosperity, And part of that is to help build those connections that you mentioned.
So build the connections between the individual and his or her financial institutions. to build connections between the individual and his or her financial decisions and give them more insight into their financial decision-making process. to help educate people about the best ways to use their finances to help power prosperity.
These are some of the areas that we're focused on, and it is a worldwide mission.
The fact is that we operate in the United States.
We have a huge customer base here, but we also operate in other parts of the world.
And as we build tools and technologies and products that can work in other parts of the world, we're hoping that those people can also enjoy the same benefits that people here enjoy.
So if I can ask you to put your own predictive hat on, so to speak, and look. five or 10 years out to the future.
Where do you see this headed? Or is there... you know, sort of a goal in the distance or something that you think is coming that's going to change or even just be a big marker in evolving the landscape for this pursuit of personal prosperity for more people?
Yeah, I think there are a couple of things that are going to happen.
I think we're going to start to see From the technology perspective, we're going to see significant acceleration in the way natural language understanding and conversational systems. and machine learning come together.
So right now, these things are related, but they're not tightly related and coupled.
So what I mean by that is that we can have chatbots and we can have predictive models and other types of models that are running, but they don't really work together. in a seamless fashion.
And those things as they come together, what I envision is a future where Machines can help people make better financial decisions based on a huge amount of data, as well as based on a huge amount of knowledge and expertise that's encoded in the system itself.
So that as people make their decisions, they can see the impact of their decisions, what's working and what's not, and it can help them make better decisions in the future.
So it's really about what we call in the field decision-making under uncertainty.
We don't know what's coming in the future, but we want to use the data and information that we have today to make the best decisions possible.
The company is Intuit. They're at Intuit.com.
The Intuit blog is at Intuitblog.com. And Ashok Srivastava, thank you so much for coming on.
It's a fascinating conversation and something that really applies to all of us in one way or another.
So thank you so much and continued luck and best success in helping everybody prosper.
Nala, thank you so much for inviting me.
Thank you. ¶¶
