The next BriefingsDirect Voice of the Customer digital transformation case study highlights how high-performing big-data analysis powers an innovative artificial intelligence (AI)-based investment opportunity and evaluation tool. We'll learn how LogitBot in New York identifies, manages, and contextually categorizes truly massive and diverse data sources.
By leveraging entity recognition APIs, LogitBot not only provides investment evaluations from across these data sets, it delivers the analysis as natural-language information directly into spreadsheets as the delivery endpoint. This is a prime example of how complex cloud-to core-to edge processes and benefits can be managed and exploited using the most responsive big-data APIs and services.
To describe how a virtual assistant for targeting investment opportunities is being supported by cloud-based big-data services, we're joined by Mutisya Ndunda, Founder and CEO of LogitBot and Michael Bishop, CTO of LogicBot, in New York. The discussion is moderated by BriefingsDirect's Dana Gardner, Principal Analyst at Interarbor Solutions.
Here are some excerpts:
Gardner: Let’s look at some of the trends driving your need to do what you're doing with AI and bots, bringing together data, and then delivering it in the format that people want most. What’s the driver in the market for doing this?
Ndunda: LogitBot is all about trying to eliminate friction between people who have very high-value jobs and some of the more mundane things that could be automated by AI.
Today, in finance, the industry, in general, searches for investment opportunities using techniques that have been around for over 30 years. What tends to happen is that the people who are doing this should be spending more time on strategic thinking, ideation, and managing risk. But without AI tools, they tend to get bogged down in the data and in the day-to-day. So, we've decided to help them tackle that problem.
Gardner: Let the machines do what the machines do best. But how do we decide where the demarcation is between what the machines do well and what the people do well, Michael?
Bishop: We believe in empowering the user and not replacing the user. So, the machine is able to go in-depth and do what a high-performing analyst or researcher would do at scale, and it does that every day, instead of once a quarter, for instance, when research analysts would revisit an equity or a sector. We can do that constantly, react to events as they happen, and replicate what a high-performing analyst is able to do.
Gardner: It’s interesting to me that you're not only taking a vast amount of data and putting it into a useful format and qualitative type, but you're delivering it in a way that’s demanded in the market, that people want and use. Tell me about this core value and then the edge value and how you came to decide on doing it the way you do?
Ndunda: It’s an evolutionary process that we've embarked on or are going through. The industry is very used to doing things in a very specific way, and AI isn't something that a lot of people are necessarily familiar within financial services. We decided to wrap it around things that are extremely intuitive to an end user who doesn't have the time to learn technology.
So, we said that we'll try to leverage as many things as possible in the back via APIs and all kinds of other things, but the delivery mechanism in the front needs to be as simple or as friction-less as possible to the end-user. That’s our core principle.
Bishop: Finance professionals generally don't like black boxes and mystery, and obviously, when you're dealing with money, you don’t want to get an answer out of a machine you can’t understand. Even though we're crunching a lot of information and making a lot of inferences, at the end of the day, they could unwind it themselves if they wanted to verify the inferences that we have made.
We're wrapping up an incredibly complicated amount of information, but it still makes sense at the end of the day. It’s still intuitive to someone. There's not a sense that this is voodoo under the covers.
Gardner: Well, let’s pause there. We'll go back to the data issues and the user-experience issues, but tell us about LogitBot. You're a startup, you're in New York, and you're focused on Wall Street. Tell us how you came to be and what you do, in a more general sense.
Ndunda: Our professional background has always been in financial services. Personally, I've spent over 15 years in financial services, and my career led me to what I'm doing today.
In the 2006-2007 timeframe, I left Merrill Lynch to join a large proprietary market-making business called Susquehanna International Group. They're one of the largest providers of liquidity around the world. Chances are whenever you buy or sell a stock, you're buying from or selling to Susquehanna or one of its competitors.
What had happened in that industry was that people were embracing technology, but it was algorithmic trading, what has become known today as high-frequency trading. At Susquehanna, we resisted that notion, because we said machines don't necessarily make decisions well, and this was before AI had been born.
Internally, we went through this period where we had a lot of discussions around, are we losing out to the competition, should we really go pure bot, more or less? Then, 2008 hit and our intuition of allowing our traders to focus on the risky things and then setting up machines to trade riskless or small orders paid off a lot for the firm; it was the best year the firm ever had, when everyone else was falling apart.
That was the first piece that got me to understand or to start thinking about how you can empower people and financial professionals to do what they really do well and then not get bogged down in the details.
Then, I joined Bloomberg and I spent five years there as the head of strategy and business development. The company has an amazing business, but it's built around the notion of static data. What had happened in that business was that, over a period of time, we began to see the marketplace valuing analytics more and more.
Make a distinction
Part of the role that I was brought in to do was to help them unwind that and decouple the two things -- to make a distinction within the company about static information versus analytical or valuable information. The trend that we saw was that hedge funds, especially the ones that were employing systematic investment strategies, were beginning to do two things, to embrace AI or technology to empower your traders and then also look deeper into analytics versus static data.
That was what brought me to LogitBot. I thought we could do it really well, because the players themselves don't have the time to do it and some of the vendors are very stuck in their traditional business models.
Bishop: We're seeing a kind of renaissance here, or we're at a pivotal moment, where we're moving away from analytics in the sense of business reporting tools or understanding yesterday. We're now able to mine data, get insightful, actionable information out of it, and then move into predictive analytics. And it's not just statistical correlations. I don’t want to offend any quants, but a lot of technology [to further analyze information] has come online recently, and more is coming online every day.
For us, Google had released TensorFlow, and that made a substantial difference in our ability to reason about natural language. Had it not been for that, it would have been very difficult one year ago.
At the moment, technology is really taking off in a lot of areas at once. That enabled us to move from static analysis of what's happened in the past and move to insightful and actionable information.
Ndunda: What Michael kind of touched on there is really important. A lot of traditional ways of looking at financial investment opportunities is to say that historically, this has happened. So, history should repeat itself. We're in markets where nothing that's happening today has really happened in the past. So, relying on a backward-looking mechanism of trying to interpret the future is kind of really dangerous, versus having a more grounded approach that can actually incorporate things that are nontraditional in many different ways.
So, unstructured data, what investors are thinking, what central bankers are saying, all of those are really important inputs, one part of any model 10 or 20 years ago. Without machine learning and some of the things that we are doing today, it’s very difficult to incorporate any of that and make sense of it in a structured way.
Gardner: So, if the goal is to make outlier events your friend and not your enemy, what data do you go to to close the gap between what's happened and what the reaction should be, and how do you best get that data and make it manageable for your AI and machine-learning capabilities to exploit?
Ndunda: Michael can probably add to this as well. We do not discriminate as far as data goes. What we like to do is have no opinion on data ahead of time. We want to get as much information as possible and then let a scientific process lead us to decide what data is actually useful for the task that we want to deploy it on.
As an example, we're very opportunistic about acquiring information about who the most important people at companies are and how they're connected to each other. Does this guy work on a board with this or how do they know each other? It may not have any application at that very moment, but over the course of time, you end up building models that are actually really interesting.
We scan over 70,000 financial news sources. We capture news information across the world. We don't necessarily use all of that information on a day-to-day basis, but at least we have it and we can decide how to use it in the future.
We also monitor anything that companies file and what management teams talk about at investor conferences or on phone conversations with investors.
Bishop: Conference calls, videos, interviews.
Audio to text
Ndunda: HPE has a really interesting technology that they have recently put out. You can transcribe audio to text, and then we can apply our text processing on top of that to understand what management is saying in a structural, machine-based way. Instead of 50 people listening to 50 conference calls you could just have a machine do it for you.
Gardner: Something we can do there that we couldn't have done before is that you can also apply something like sentiment analysis, which you couldn’t have done if it was a document, and that can be very valuable.
Bishop: Yes, even tonal analysis. There are a few theories on that, that may or may not pan out, but there are studies around tone and cadence. We're looking at it and we will see if it actually pans out.
Gardner: And so do you put this all into your own on-premises data-center warehouse or do you take advantage of cloud in a variety of different means by which to corral and then analyze this data? How do you take this fire hose and make it manageable?
On the Google side, we take advantage of Bigtable and BigQuery and some of their infrastructure tools. And we have good, old PostgreSQL in there, as well as DataStax, Cassandra, and their Graph as the graph engine. We make liberal use of HPE Haven APIs as well and TensorFlow, as I mentioned before. So, it’s a smorgasbord of things you need to corral in order to get the job done. We found it very hard to find all of that wrapped in a bow with one provider.
We're big proponents of Kubernetes and Docker as well, and we leverage that to avoid lock-in where we can. Our workload can migrate between Google and the SoftLayer Kubernetes cluster. So, we can migrate between hardware or virtual machines (VMs), depending on the horsepower that’s needed at the moment. That's how we handle it.
Gardner: So, maybe 10 years ago you would have been in a systems-integration capacity, but now you're in a services-integration capacity. You're doing some very powerful things at a clip and probably at a cost that would have been impossible before.
Bishop: I certainly remember placing an order for a server, waiting six months, and then setting up the RAID drives. It's amazing that you can just flick a switch and you get a very high-powered machine that would have taken six months to order previously. In Google, you spin up a VM in seconds. Again, that's of a horsepower that would have taken six months to get.
Gardner: So, unprecedented innovation is now at our fingertips when it comes to the IT side of things, unprecedented machine intelligence, now that the algorithms and APIs are driving the opportunity to take advantage of that data.
Let's go back to thinking about what you're outputting and who uses that. Is the investment result that you're generating something that goes to a retail type of investor? Is this something you're selling to investment houses or a still undetermined market? How do you bring this to market?
Natural language interface
Ndunda: Roboto, which is the natural-language interface into our analytical tools, can be custom tailored to respond, based on the user's level of financial sophistication.
At present, we're trying them out on a semiprofessional investment platform, where people are professional traders, but not part of a major brokerage house. They obviously want to get trade ideas, they want to do analytics, and they're a little bit more sophisticated than people who are looking at investments for their retirement account. Rob can be tailored for that specific use case.
He can also respond to somebody who is managing a portfolio at a hedge fund. The level of depth that he needs to consider is the only differential between those two things.
In the back, he may do an extra five steps if the person asking the question worked at a hedge fund, versus if the person was just asking about why is Apple up today. If you're a retail investor, you don’t want to do a lot of in-depth analysis.
Bishop: You couldn’t take the app and do anything with it or understand it.
Ndunda: Rob is an interface, but the analytics are available via multiple venues. So, you can access the same analytics via an API, a chat interface, the web, or a feed that streams into you. It just depends on how your systems are set up within your organization. But, the data always will be available to you.
Gardner: Going out to that edge equation, that user experience, we've talked about how you deliver this to the endpoints, customary spreadsheets, cells, pivots, whatever. But it also sounds like you are going toward more natural language, so that you could query, rather than a deep SQL environment, like what we get with a Siri or the Amazon Echo. Is that where we're heading?
Bishop: When we started this, trying to parameterize everything that you could ask into enough checkboxes and forums pollutes the screen. The system has access to an enormous amount of data that you can't create a parameterized screen for. We found it was a bit of a breakthrough when we were able to start using natural language.
TensorFlow made a huge difference here in natural language understanding, understanding the intent of the questioner, and being able to parameterize a query from that. If our initial findings here pan out or continue to pan out, it's going to be a very powerful interface.
I can't imagine having to go back to a SQL query if you're able to do it natural language, and it really pans out this time, because we’ve had a few turns of the handle of alleged natural-language querying.
Gardner: And always a moving target. Tell us specifically about SentryWatch and Precog. How do these shake out in terms of your go-to-market strategy?
How everything relates
Ndunda: One of the things that we have to do to be able to answer a lot of questions that our customers may have is to monitor financial markets and what's impacting them on a continuous basis. SentryWatch is literally a byproduct of that process where, because we're monitoring over 70,000 financial news sources, we're analyzing the sentiment, we're doing deep text analysis on it, we're identifying entities and how they're related to each other, in all of these news events, and we're sticking that into a knowledge graph of how everything relates to everything else.
It ends up being a really valuable tool, not only for us, but for other people, because while we're building models. there are also a lot of hedge funds that have proprietary models or proprietary processes that could benefit from that very same organized relational data store of news. That's what SentryWatch is and that's how it's evolved. It started off with something that we were doing as an import and it's actually now a valuable output or a standalone product.
Precog is a way for us to showcase the ability of a machine to be predictive and not be backward looking. Again, when people are making investment decisions or allocation of capital across different investment opportunities, you really care about your forward return on your investments. If I invested a dollar today, am I likely to make 20 cents in profit tomorrow or 30 cents in profit tomorrow?
We're using pretty sophisticated machine-learning models that can take into account unstructured data sources as part of the modeling process. That will give you these forward expectations about stock returns in a very easy-to-use format, where you don't need to have a PhD in physics or mathematics.
You just ask, "What is the likely return of Apple over the next six months," taking into account what's going on in the economy. Apple was fined $14 billion. That can be quickly added into a model and reflect a new view in a matter of seconds versus sitting down in a spreadsheet and trying to figure out how it all works out.
Gardner: Even for Apple, that's a chunk of change.
Bishop: It's a lot money, and you can imagine that there were quite a few analysts on Wall Street in Excel, updating their models around this so that they could have an answer by the end of the day, where we already had an answer.
Gardner: How do the HPE Haven OnDemand APIs help the Precog when it comes to deciding those sources, getting them in the right format, so that you can exploit?
Ndunda: The beauty of the platform is that it simplifies a lot of development processes that an organization of our size would have to take on themselves.
The nice thing about it is that a drag-and-drop interface is really intuitive; you don't need to be specialized in Java, Python, or whatever it is. You can set up your intent in a graphical way, and then test it out, build it, and expand it as you go along. The Lego-block structure is really useful, because if you want to try things out, it's drag and drop, connect the dots, and then see what you get on the other end.
For us, that's an innovation that we haven't seen with anybody else in the marketplace and it cuts development time for us significantly.
Gardner: Michael, anything more to add on how this makes your life a little easier?
Bishop: For us, lowering the cost in time to run an experiment is very important when you're running a lot of experiments, and the Combinations product enables us to run a lot of varied experiments using a variety of the HPE Haven APIs in different combinations very quickly. You're able to get your development time down from a week, two weeks, whatever it is to wire up an API to assist them.
In the same amount of time, you're able to wire the initial connection and then you have access to pretty much everything in Haven. You turn it over to either a business user, a data scientist, or a machine-learning person, and they can drag and drop the connectors themselves. It makes my life easier and it makes the developers’ lives easier because it gets back time for us.
Gardner: So, not only have we been able to democratize the querying, moving from SQL to natural language, for example, but we’re also democratizing the choice on sources and combinations of sources in real time, more or less for different types of analyses, not just the query, but the actual source of the data.
Ndunda: Again, the power of a lot of this stuff is in the unstructured world, because valuable information typically tends to be hidden in documents. In the past, you'd have to have a team of people to scour through text, extract what they thought was valuable, and summarize it for you. You could miss out on 90 percent of the other valuable stuff that's in the document.
With this ability now to drag and drop and then go through a document in five different iterations by just tweaking, a parameter is really useful.
Gardner: So those will be IDOL-backed APIs that you are referring to.
Bishop: It’s something that would be hard for an investment bank, even a few years ago, to process. Everyone is on the same playing field here or starting from the same base, but dealing with unstructured data has been traditionally a very difficult problem. You have a lot technologies coming online as APIs; at the same time, they're also coming out as traditional on-premises [software and appliance] solutions.
We're all starting from the same gate here. Some folks are little ahead, but I'd say that Facebook is further ahead than an investment bank in their ability to reason over unstructured data. In our world, I feel like we're starting basically at the same place that Goldman or Morgan would be.
Gardner: It's a very interesting reset that we’re going through. It's also interesting that we talked earlier about the divide between where the machine and the individual knowledge worker begins or ends, and that's going to be a moving target. Do you have any sense of how that changes its characterization of what the right combination is of machine intelligence and the best of human intelligence?
Ndunda: I don’t foresee machines replacing humans, per se. I see them empowering humans, and to the extent that your role is not completely based on a task, if it's based on something where you actually manage a process that goes from one end to another, those particular positions will be there, and the machines will free our people to focus on that.
But, in the case where you have somebody who is really responsible for something that can be automated, then obviously that will go away. Machines don't eat, they don’t need to take vacation, and if it’s a task where you don't need to reason about it, obviously you can have a computer do it.
What we're seeing now is that if you have a machine sitting side by side with a human, and the machine can pick up on how the human reasons with some of the new technologies, then the machine can do a lot of the grunt work, and I think that’s the future of all of this stuff.
Bishop: What we're delivering is that we distill a lot of information, so that a knowledge worker or decision-maker can make an informed decision, instead of watching CNBC and being a single-source reader. We can go out and scour the best of all the information, distill it down, and present it, and they can choose to act on it.
Our goal here is not to make the next jump and make the decision. Our job is to present the information to a decision-maker.
Gardner: It certainly seems to me that the organization, big or small, retail or commercial, can make the best use of this technology. Machine learning, in the end, will win.
Ndunda: Absolutely. It is a transformational technology, because for the first time in a really long time, the reasoning piece of it is within grasp of machines. These machines can operate in the gray area, which is where the world lives.
Gardner: And that gray area can almost have unlimited variables applied to it.
Ndunda: Exactly. Correct.