If you would like to earn CPE credit for listening to the show, visit earmarkcpe.com slash FPA.
Download the app, take a short quiz, and get your CPE certificate.
Finally, if you enjoy listening to FPNA today, please go to your podcast platform of choice, click the subscribe button, and leave a rating and review of the show.
And now, onto the show. From Data Reels, this is FPNA Today.
Welcome to FPNA Today. I'm your host, Glenn Hopper.
Our guest today is Brandon Wilson, a digital transformation pioneer with a comprehensive background in customer software solutions, automation, and AI.
Brandon has driven multi-million dollar initiatives in both startup and enterprise environments, mastering innovation in the digital era.
He excels in sales and business development, establishing platforms that address challenges in digital product development, process automation, AI, and large language models.
His specialties include strategic planning and analysis, new business development, key account management, digital transformation, product management, go-to-market strategy, project governance, client relationship building, and team leadership.
As the solution architect, Brandon focuses on embedded systems, IoT, AI, slash LLMs, and Web3, empowering individuals in software and technology.
He collaborates with clients to create digital solutions and conducts competitive analysis to ensure market positioning.
Brandon leads cross-functional teams, ensuring synergy and exceptional results, and maintains strong relations with key clients and stakeholders through excellent communication skills.
We're excited to have Brandon Wilson on the show today to share his insights and experiences in driving digital transformation and innovation.
Welcome, Brandon. Brandon Wilson, thank you.
Glad to be here. Appreciate it.
Brandon Wilson, thank you.
I know we've got a lot to cover today, and super excited.
One of the things I've loved about this show is the number of data geeks.
I've gotten to talk to lately, and I'm putting you in that bunch, and I think that's probably a moniker you would proudly wear.
Yes, all accept me. Glad to.
I guess just as a little background, give me a bit.
I mean, I went through kind of your professional bio, but tell me about your career journey and what led you to your current role at steady dynamic.
Yeah, sure. So I started my career, and most of my career was spent in telecom equipment, and I had the opportunity to lead an initiative for the introduction of what was a very early version of smart home sensors and devices.
And this was done for the cable television industry through a platform called iControl.
And although the sensors at the time were simple, they were really the first, I guess, broad use of what I'll call smart home features.
And so this is where I kind of caught the bug of how software going from unintelligent to intelligent devices and how software and then data collection from these sensors could actually automate, trigger, and enrich an experience for instance, and smart home applications.
And so that's where I heard that was probably good 15, 18 years ago at this point, but that's really where I first caught the bug and started to see the importance of not just intelligent or smart devices, but then leveraging data from those devices for a number of beneficial outcomes.
Gotcha. And there's something about telecom, and I think it's because there's so much data.
So my start was in telecom as well.
And it's really, I think that's where I got my love for data and realizing how you could integrate data from different sources and put it into financial planning and operational planning and all that.
So it's, yeah, I've talked to several people in telecom had a few guests on the show from that background and to a person, they were a very data centric and you know, you can see from the telecom start how that how they help me they sit on significant pipelines for data collection if you think about it.
And so it makes sense to their their that industry to be kind of the pioneers in that space.
Yeah, and it's been for me, it's been two decades since I've been in telecom and I haven't been at another company that had the level of data that we did at telecom.
So I missed those days where I had, you know, just a full universe of data to play with like that.
I can remember watching the progress of AT&T's Pleasanton, California Data Center and watching how much it grew or the accommodation of the vast amounts of data that they were collecting.
And it was it was a wild ride to watch from its very early days of what it was and how it grew and how quickly it grew and and you know, there's just there's a again, just a ton of data collection out there.
And I guess we know part of today's discussion is really being productive with that data.
How do you use it? You know, how what are the what are the pros, the cons, the risks, the rewards, and how do you really bake it in to doing more advanced analytics as it might relate to, you know, financial planning and forecasting and business health management.
It's it's a it's a really exciting, you know, time to be in this space because there's a lot of great progress.
It is and we're you know, we're about to shift it up even to another level as the proliferation of data and the availability and the amount of data that's out there has grown.
But now with all these super computers being built around to power the training of these new LOMs and frontier models that are out there, I mean, we're about to see the, you know, next level of the data.
So like when you and I started our career in telecom and watching that, you're seeing the next level of it.
So now you're the founder and CEO of Steady Dynamic.
Tell me about Steady Dynamic, what you guys do and some of the key services and solutions you offer.
Sure. Yeah. So at our core, we are a custom software development, consultancy, and agency.
So we build the spoke custom software applications for our clients that in large part includes, you know, traditional web and mobile applications.
But for the better part of the last say two years has seen a significant uptick in the implementation integration of AI and probably majority of those applications and use cases are for internal business process either the promotion of efficiency or enrichment.
So thinking, you know, kind of right data in the right hands at the right time for the right reason.
And so it's been really interesting to watch.
I think we're probably at coming out of kind of phase two of what I'll call the kind of AI maturity model where a lot of our clients have done the homework, maybe even some experimentation.
So they recognize there's value there.
But I think it's in that same process of experimentation and diligence that they've also recognized the limitations and that and that they there's a need to call in the pros to help them move to what I'll call more production environments where not only are they getting the results that are expected specifically from the AI engine portion of the application, but also are being
mindful of, you know, security and governance that go along with it.
And so we're seeing a dramatic uptick in in in I think that client persona that has done the experimentation knows there's something there is prepared to invest, but is now is now requiring kind of outside help to make sure that you know, there's a better chance at harvesting ROI.
But again, really built in in a production way that is secure and safe to roll out for both internal and external use cases.
Yeah. And one of the things when you and I were talking before the show, I realized we're both going after the same goal, but we're coming at it, you know, you're coming into the house through the door and I'm climbing in through the window as the finance guy in that we're both trying to get access to the data to make the business run better.
And with what you guys do, you know, it's on process and operations and automation and using data to to drive decisions and build out what you can do with the data, the amount of data you're collecting and everything.
And for me, I want the same thing, but my initial desire for this data was because I thought I can use this to help drive my models and help, you know, give me more data and I can make a more accurate model and better predictions.
But we're going after the same thing and I know, you know, I think when you when you engage with the client, probably the CFO is not the first call that you have with with the client.
But I do think that it is there's crossover in all this data.
So I'm wondering, you know, as you you come into a business and you're seeing how they use the data and now understanding where how finance can use it and you talked about as being at this sort of next level of AI maturity, how do you see as as we get more mature with our data, how can you see AI, generative AI and their traditional machine learning and AI that's been out
there? How do you see it further transforming the finance industry?
And I'm thinking particularly in finance forecasting, but I know it has other applications as well.
Yeah, sure. So, so, you know, I think for the most part, most people, let's say in the FPNA industry have to some degree implemented what I'll call, I guess, you know, predictive analytics, whether that's either custom or bespoke modeling or just using tools that come in SaaS platforms that they're using today and have a pretty good handle on how that applies.
I think the next generation and it is related to a number of advancements, I think in the industry from a technology perspective are the unlocking of prescriptive analytics.
And so, couple not just data presentation or data for the purpose of kind of extracting insights, but now insights that come with recommendations and the ability to run multiple scenarios, even thinking of it as kind of a digital twin.
So, you could, you know, simulate or create millions of simulations based on whatever inputs you want to change in this in order to gain prescriptive analytics.
And then I think if you couple generative AI into that equation, can also, you can harvest the best of the best some of the best features of generative AI in that it is natural language in order to create reports and insights for different stakeholders in the FPNA industry and because they all might have kind of different requirements or different views, so to speak.
And so, the rapid conversion of those prescriptive analytics, so say for report generated for the accounting department or the CFO and or the business unit leader, I think is a really exciting opportunity because it makes the availability of the data and insights they're more consumable in a sense and more applicable to the stakeholder involved and what they're
really trying to gain from that data.
And I think probably the other piece of the puzzle is getting to more real-time analytics.
And so, you know, traditional cadence of, you know, an FPNA professional might be to do weekly and monthly and quarterly and really and a lot of that, a lot, sometimes that work is very kind of forklift or kind of waterfall.
You know, you gear up a bunch of effort, you produce a document and then, you know, you move on to the next kind of checkpoint in the sequence, but I think real-time data acquisition and analytics is also empowering us to do things more on a daily, if not hourly basis and see things way sooner from from an analytics and forecasting perspective.
Yeah, and that real-time part is so key because I can remember, you know, back, back when I started where everything you're waiting on some crime job to run overnight and you're just waiting for reports and everything.
And now, as you shift to more and more live data and being able to use it, and you see companies going to, you know, real-time close kind of they'll always be closing where you're not just waiting for the end of the month.
Now, there's the official financial close that you still have to go through, but having access to all this data as you go along.
And another thing that really, I think for finance people, when we think about analytics, you know, I think about the level of analytics you can do and sort of the data maturity of a company where the first thing that companies do and hopefully in 2024, you know, we're all there now, but I know there are still companies that are struggling here, but just that that descriptive
analytics. So just this is the universe of data we have.
What does it mean? What do I know about my customers?
What do I know about customers who've churned about our sales process, our pipeline, you know, just taking all the taking stock of all your data and saying, here's the charts and graphs that show kind of where we are today.
And then evolving that into, okay, based on this, I can now apply that to my modeling and I can have predictive analytics, which is great because it lets you, you know, come up with those more accurate forecast based on more data.
But I think that the way that this resonates with finance people is to move from that sort of cost center label that we get settled with where we're considered backward looking.
And even if we build good models, it's like, yeah, that's fine.
You're modeling out a potential future.
But how can I turn that into strategy?
And you mentioned going to prescriptive and for finance people, that's here's our data.
Here's the KPIs we're tracking.
But now I've found the levers.
If you want to change the future, pull this lever and all that.
So being able to identify that is that next level where, oh, now we're suddenly a strategic partner.
And I think AI and data is the way that we're going to really drive that home and increase our value in the company.
Couldn't agree more. And taking that approach, obviously, and to your point about kind of changing it from a cost to a value driver in that respect is exactly the kind of the point of what you're going to implement those type of strategies for is, you know, if you think about just principally the impact that something like one of those levers being pulled from in cash
flow analysis and cash acceleration, I mean, these are really important variables in the health management of a business.
And the larger you get the more important, impactful and important, some of those decisions are.
And so I couldn't agree more.
I think that's, and I like saying it, like, to move from the kind of the cost to the value add model by being able to know which levers to pull is a great way to look at it.
Yeah, and I think for a lot of our audience, certainly finance professionals like everyone else, we're getting more data savvy and we understand the value of it and how to use it more and more.
But I think what a lot of us are challenged with is we know data is out there.
We're not the gate keepers of the data.
We know there are certain data points in our CRMs and our ERP systems, maybe in project management tools, there's these data points that are out there.
If we could get them or understand them, and you know, if you've got a company that has a really well-defined data warehouse where you understand, you know, the source of truth and everything and what they all mean, that's huge.
But thinking of we have this data and these various sources, how do you see once we solve for the delivery of it, how can that be integrated to kind of enhance financial functions?
If I know, you know, something about pipeline, maybe there's this mentality of, well, that's in my CRM tool, but we never won the deal.
So is that data valuable to me?
Or, you know, maybe different points that we don't think about.
Can you think of some examples of how that external data might be able to be integrated to enhance modeling and other financial functions?
Yeah, sure. I mean, I think first and foremost, if you think of all the what might be viewed as disparate systems, right?
So CRM, ERP, you might have project management tools, you know, a number of things that are used to harvest data or at least have the ability to harvest data.
If you think about them all as opposed to being disparate tools as interconnected tools, they allow just for first and foremost, a more comprehensive kind of 360 view of your total business.
And then, you know, maybe the example one that you started there is if you think, well, I guess I'll kind of stick to cash to book as maybe an example here, which even though you might not be able to do proper revenue recognition until the order is actually booked, you can study sentiment analysis or, you know, behave your patterns from your CRM tool about like cadence
of inter, you know, intertouch points between a client as you're going through a sales funnel to start looking at maybe more prediction toward whether that client is going to convert.
And so while that may not be traditionally viewed as revenue recognition, it's a leading indicator that revenue is going to be booked, right?
So you can get even further ahead of the actual revenue recognition event.
And then on the other side, as an example, you could use, you know, various clustering models, ones that are just kind of widely available and known to analyze, you know, your expected payment terms.
How much cash is going to be out according to, for instance, the different sizes of companies, the different industries that they represent, the different products that they buy and really get prediction models around, you know, okay, their terms are net 30, but when do you actually expect to be paid?
And so you're actually kind of lengthening the overall, let's say pipeline of analysis to maybe enable you to do better, just, you know, kind of threshold analysis.
So you have a low watermark expected and a high watermark.
And if you start blending in some of these other components that come from these other tools, you might gain, you know, not only better analytics, but going back to now starting to look at the levers you pull, you know, maybe you can change those dynamics.
Is there something you can do to accelerate from a CRM perspective in that sales pipeline?
You can accelerate that conversion and or on the other end, you know, being able to collect the cash payment quicker in the cycle based on your understanding or expectations that the analysis provides.
Yeah, and I love that because I think of, you know, it's funny because in recent years, I feel like sales and marketing has really, they've gotten deep into data analytics and they're using these tools a lot more.
And I think finance and FBNA are catching up now, but sales and marketing really was on the leading edge with using machine learning and using customer data and using all this to to do their forecasts.
But as we look at what data we have in the different systems, and you know, whether it's predicting customer churn or understanding customer acquisition cost or lifetime value, I mean, it all it all goes in and feeds the model and the financials.
So it's we're beyond just analyzing our financial statements.
It's again going to those levers of what impact.
So this is our pipeline.
This is what we won. This is what we didn't.
What can how can we predict to make better forecasting models?
It's we need to know the data about as much data about the ones who didn't sign up as the ones who did.
So it's sort of looking at the broader, you know, beyond just the the GL and the other parts of the business to do our jobs.
Another thing in the data analysis is say we have say our company has reached a level of data maturity where we understand, you know, we're all we have known KPIs, we have known sources of truth, we understand the data and disparate systems and what to use.
But I think for a lot of finance people, it's, you know, we build models out in Excel and we kind of, okay, I've got this data and I've got this data.
Let me put it in and see how it informs the model.
But I'm wondering, you know, one of the other things you and I talked about before the show is is moving kind of beyond Excel and moving into whether it's our Python or something where you kind of programmatic languages.
But tell me some of your experience and I think this would translate into finance, but we're using these more complex models to find, find correlations.
Maybe that aren't readily apparent and that you wouldn't even, you know, if you weren't doing a correlation matrix, you wouldn't even know these two things are correlated.
But some of the complex models that you use to sort of make sense of this big data where you take your internal data sources that you have and maybe you start looping in external data sources macroeconomic information.
Do you have some examples of that?
Yeah, sure. And I think although it's an obvious statement, this is really the the expansiveness of data availability is really what drives the need for the models in the first place because that's just simply not possible to do as a human.
I'm not sure if you remember seeing the movie, the accountant with Ben Affleck and the Romanies, writing everything he's pouring through papers and writing, so okay, not all of us are subordinates.
So this is why we use machine learning and the good news is that there's a lot of off-the-shelf models that have been perfected that are useful as it might apply to finance, profession, and industry.
So, you know, you know, with the the tried-true kind of OG linear regression, which, you know, is still, you know, what I'll call it, a fairly simpler model that you use, but, you know, relying on historical data to project the future, but it's still important from a perspective of vast amounts of data, right?
But I'm a big fan of clustering, so like nearest neighbor models, which might reveal insights based on other, for instance, if you're trying to look at other either transactions, other entities, kind of other scenarios, how they relate to other knowns in the system to be able to to drive correlations that way.
And then another example is using naive bays, which is is really starting to get cool into probabilistic outcomes based on prediction.
And again, these only were if they're large data sets.
And so, you would implement them only if you have significant data availability, in which case, to kind of flip from, you know, going from, you know, human heuristics into proper machine learning and using these models.
But I mean, like I said, the really good news is that a lot of the established available machine learning models, while they may not have been necessarily designed specifically for a financial analysis directly apply, but it's always about picking the right model for the right reason.
Yeah. And I think, you know, to your point on you have to have the data to do anything with it, I think about companies that have had great success.
And, you know, using machine learning for more than a decade now.
So e-commerce companies have so much data and they're able to do, they can do their clustering and segmentation and building out the models based on that.
But if you're, you know, you're a chain of dry cleaners that have 20 locations, you know, what data do you have that you're you're able to use.
So it's, it is a challenge.
So I guess there's on the low end, there's people who just don't have enough internal data.
So how could you use AI to improve your business?
And that's a challenge.
But and then even on the other side of it, there's sort of too much information trying to figure out what of it you can you can use.
So from your viewpoint, and as you go in and work with clients, what are some common challenges that companies face in accessing data and making use of it?
And how do you help them address some of these challenges?
Yeah, sure. It's it is probably still remains the largest challenge to successfully implementing really any analytics or AI strategy is the usefulness of the data.
It was interesting. I've been prepped for the show.
I've known these numbers or I've known at least anecdotally the significance of some of these numbers.
But I actually looked it up and there's a recent report from NTT data that said 80% of the world's data.
So this is all data availability is unstructured and 90% of that unstructured data.
So 72% of all data is not used in any form of analysis.
And it begs the question why?
Well, sometimes the answer I think is people don't really know how to use the data as an example.
I think people over collect and store data maybe in some cases unnecessarily.
I had a recent conversation with someone and they were contemplating the elimination of some historical data.
And I asked like, you know, sit down with your team and think about if you can determine any usefulness to it, keep it.
If you can't get rid of it because you're paying for it.
You're paying to store data.
You're just creating reams and reams of data that you're not using.
So, but at a structural level, you know, it's important to note that it can be expensive to convert unstructured data to structured data environments.
So getting from data lakes to data warehouses.
And while I think that's the right approach to implement if you're really locked down in terms of the purposefulness of your kind of AI or email implementation.
So meaning, you know exactly what you're looking for.
You know exactly the data that's going to support it.
In which case you want to construct, you know, just just rock saw would lock down repeatable and scalable models.
But what's been interesting here more recently, I think, is the proliferation and advancement of tools that allow you to interact with unstructured data.
And so, while that may not be the production choice, I think it provides an opportunity that historically maybe wasn't always necessarily available or even practical has allowed people to experiment with the unstructured data in ways that that, you know, okay, it's not that recent.
But let's say generally speaking in data science is more a more recent phenomena.
And that that can be through tools like, you know, in Alterics, which helps you, you know, kind of, I wouldn't say autonomously, but at least assisted unstructured data preparation for analysis.
You get, you know, obviously huge platforms like Databricks, which foundationally are data lake, not data warehouse, so to speak.
There's talent, which is a great tool for organizing disparate sources.
So there's been a lot of progress in terms of the tools that help account for, let's say some of the challenges that are related to making data useful.
And then not the least of which is what we're seeing with the LLMS.
And some of these foundational models that it doesn't care whether it's structured unstructured, you know, multimodal or otherwise, right?
So it's going to take a look at just kind of every every data source you've got and do its best with what you put in the engine.
So yeah, I think it will always remain a challenge.
And there's a reason why from that statistic so much is unused, but I think generally speaking, the tools that are technology, I should say, that is more recently been introduced to the industry is making it easier to put that structured and unavailable data to wear.
FPNA today is brought to you by Data Rails, the world's number one FPNA solution.
Data Rails is the artificial intelligence-powered financial planning and analysis platform built for Excel users.
That's right, you can stay in Excel.
But instead of facing hell for every budget, month and close or forecast, you can enjoy a paradise of data consolidation, advanced visualization, reporting and AI capabilities.
Plus, game-changing insights giving you instant answers and your story created in seconds.
Find out why more than a thousand finance teams use Data Rails to uncover their company's real story.
Don't replace Excel, embrace Excel.
Learn more at datarails.com.
Just thinking about where we are with LLMs right now and being able to take these highly trained generalists and be able to apply them to our own data sets and be able, you know, we talked about the limitations of retrieval augmented generation and how companies can actually use generative AI to really focus on their data and turn those generalists into very specialized
tools to use for their company.
So I think about sort of historically and where I got my start, where I think and I was super smart for being able to write SQL queries and be able to get access to this data.
But it's grown so much into your point on data not being used.
So much of the data out there is unstructured now and I'm thinking about, you know, the no-sequel sort of structure.
And even in maybe if you could touch on internet of things, the amount of data that's out there that's not being used.
I mean, that's, there's a treasure trove out there, but it's so much in just trying to figure out what's useful with it.
But one of the things you and I talked about was using graph databases and finance operations and the potential there.
So throwing a lot at you, but talk to me about the advantages of, you know, we have all the structured, unstructured data, internet of things, putting it into a graph database, and then what the advantages of graph are compared to maybe traditional indexing and just being able to try to make sense of all this.
If you could walk me through that and with keeping in mind that our audiences that we're going to, they're going to be switching off here if we get too far deep into the the technicalities of the data side of it, but thinking of it, the advantages of that four finance operations.
Yeah, sure. So I think first it's helpful to just kind of give an oversimplified comparison of SQL versus no SQL.
So if you think about a traditional SQL, use of traditional SQL queries, I would say that has been largely used when again, you know exactly what you're trying to achieve.
You know exactly the data source, the data type, it's, it's, it's cleanse data, kind of the perfect implementation, right?
So you know exactly what the inputs are, it's exactly what the outputs are, and that's probably historically been the primary use for SQL, whereas no SQL, you get a lot more flexibility with things like the schema, right?
That can change over time in the SQL.
And so I would just say it's somewhat inappropriate to look at it this way, but I think no SQL allows you to do more experimentation because it has less limitations of structure.
And then graph databases are actually a derivative of no SQL.
So, so it really fits into that space.
And I think the best way to think about a graph database is that it is by design intended to analyze the relationships between data points, which are called nodes in a graph database.
So it's mostly concerned with the relationship between two data points.
As opposed to the data point, it's self-souder space.
So what's really great though about a graph database is it's not bound by the rules of SQL of joiners, right?
So you don't have to do all these complex join functions that can get really complicated in out of hand.
And if you pass one thing up, it kind of defeats the whole purpose of the entire, you know, kind of the analytics or models.
So, but I think what graph database in the case of like in the financial and analytics industry is really important because first and foremost, I think like Power BI and Tableau allow you to visualize data, seeing a graph of your data is impactful by itself, specifically when you think about complexity and dependencies.
And so, as an example, if you manage a company that has multiple subsidiaries and then multiple domestications of incorporation and, you know, a number of other complexities, it can be really interesting just to see the entire graph of how your company looks and how the relationships exist, specifically, you have to deal with things like, you know, intercompany
transactions or transfer pricing, you know, things like that.
So you can really help get a visual impression, which is the start of the analysis.
But then the relationships and the semantic relationship analysis of those data points is really what makes a graph database powerful.
So as an example, and probably most people watching this were familiar with what's it called the Panama papers, where they actually reconstructed all the complexity of all those offshore accounts in a graph database and how they got to it was actually looking at the transactions themselves.
So if you consider, for instance, data that would be used in a graph database, so you start looking at like the transaction itself and then you start adding metadata to it and its relationships to say geography or entities or, you know, the different banks that were involved, the different, let's say, more, a firmographic information, like the names that were used,
the passwords, the email addresses, you know, things like that, it's what allowed them to construct the graph of how all those transactions ultimately, you know, fraudulent transactions effectively were moving through the system.
And so the use case of graph database is currently really common for actually looking at fraud, laundering, things related to, you know, KYC and Identity Management.
And so it's a really powerful tool, it's maybe a little bit lesser known in the grand scheme of things, but again, if you're trying to look at how data impacts other data, so that's a magic relationship between them, that's really where you can get a lot of great value out of using a graph database versus a traditional relational database.
Yeah, and you know, for anyone who hasn't seen them, to me, these network diagrams from graph databases are visual and they are telling and you get, it's like looking at any other chart of graph when you start seeing all the nodes and then the vectors coming off of them, connecting to other areas and you start seeing the relationship and how connected things are,
it helps you to be able to visualize and understand how, if you were querying along that path, how you might be able to access the data versus, you know, if you're looking at a traditional SQL database schema, there's nothing visual about that, but seeing it in the networks.
And I think the reason you and I were talking about it and the reason that it's important now is if we're trying to take generative AI and use it on our data, that network, those connections, that graph database, lets the LLM know where to go look for something specific, whereas if you just don't a bunch of, you know, you have a bunch of documents that you put in a vector database
and the LLM is going through and searching, it's got a limited context window, it's going to go through and, you know, may peter out, may not get what you result, but if you can start it down the path of following these graphs and then you could have other indexes in there as well, that really it's, it's like a map for how to get to your data.
And to me right now, that is very important to be able to use some of this data that we're not able to tap into right now and use the benefits of AI to do that to find these connections that we on our own couldn't.
Yeah, well said, and they took two parts, one, I'm still a huge fan of just generally speaking, what visualization does to improve your kind of thought process related to analytics.
And so, to your point of being able to see it, the other great thing that it does is allows you to see the change.
So as an example, if you're doing scenario generation or simulations with a graph database, and I'm just, since I said it, transfer pricing and you say, well, there's a transfer pricing rule change, I'm going to implement the change in the scenario and now see how it impacts the entities, it can represent that data like that.
So it'll actually change the graph to match the variable changes, which I think is just really powerful and also fun.
Yeah, I just said that was fun, but yes, that's the first time transfer pricing and fun have ever been.
All right, fair enough.
Well, we can start this podcast by saying we were data nerds.
So, but on your other point too, you're absolutely right.
There's a new, I keep saying everything's new, but I think it's just relatively new, but approach to graph rag, which is what you describe.
And I think it is a powerful addition and the arsenal to not only making data more interactive, but making the output of let's say the conversation you're having with these LLMs for the purpose of analysis to be more precise, to be more impactful and even more insightful.
And so if you think about as an example, your ability to say, well, I'd like to understand all relationships to this courtesy or this connection point between nodes and two nodes in a graph.
And then you look for the commonality across that type of relationship on a massive data set.
I mean, you can do those kind of queries with natural language using an LLM and graph-rad versus let's say traditional rag to your point, which is is picking up a level of context that might not have been present in traditional or traditional vector database and rack strategy.
So yeah, I think it's an improvement exactly as you describe, which is the contextual understanding of the data as it relates to itself and one another as opposed to just kind of a giant static data set.
Yeah, I'm going to take it all the way back and say, you know, imagine I'm an FPNA person and I do, and we still use Excel.
I mean, we love to bash it, but Excel is the first place I go to do any sort of analysis.
It's just the easiest way to access.
And I think for finance people, that's always where we're going to go.
But I think about we dove straight into the deep end, but we, if we back up to a higher level, we have access to things right now as generative AI, let's us use natural language, or just interact with our systems the same way we would interact with another human being and not have to translate into Python or horror, whatever we're doing.
But I'm trying to picture now, a lot of our listeners may be here in what we're saying and think, yeah, that's really great.
I don't have access to that.
What does this mean to me?
So I'm thinking about we see the very near future where I've just written about Lamas new, agentic rag that they're come out with that model and some different applications of that.
And so all these things we're talking about in the abstract, but to try to take them for our users and help them understand what they're going to be.
So using LLMs, using these LLM powered chatbots, AI, we've got democratization of data.
We may have increased access to data that we didn't before.
But of course, in having that, we have to have this data science understanding that we didn't before to some extent, because it's kind of like if you're not a finance person, you have to know the difference between EBITDA and that income to really have value there.
So you've got to know the terms.
You may not have to be able to be the machine learning engineer that knows how to write the code to get to them.
But we're talking about all this from a super data geek perspective.
But for frontline FPNA people right now who say, you know, I'm not going to go learn Python and all this.
But how do you see the next, I don't know how long it, every time I make a prediction, I'm wildly wrong.
So I need to get out of the prediction game.
But what does the data environment look like for say the next two, three, five years for finance professionals, and how are they going to be able to leverage it?
So all this stuff that we're talking about, where we're getting deep in the guts, where we're talking about, you know, graph databases and all these connections and being able to access the data.
But for end users of this data, what does the future look like for them?
And how are they going to be able to take advantage of this kind of technology?
Yeah, well, I think first and foremost, it starts with the availability of tools, not the least of which is the LLMs and the implementation of things like chat, GPT or G, I should say GPT4 and the LOMMA models is to allow you to interact with your data without being a data sign.
So first and foremost, that's obviously one of the greatest advantages.
And I think the only limitation at this point, assuming you have data and interact with, is curiosity.
And so the first and probably best thing that has happened with the progress with LLMs is the ability to just obviously use natural language processing, natural language to have conversations with your data.
And I've experimented with this extensively.
You should assume it knows everything about data science that needs to be known.
And you can tell it I'm not a data scientist and instead have just a curiosity conversation, the what ifs, you know, go back to very some, back this, maybe some very simple Excel formula ways of thinking.
You don't need to worry about the complex models, at least certainly not to start, right?
So great, it's a great pathway, maybe to get to more advanced data.
And I think the future data is both a blessing and a curse.
And I think one of the things that's going to impact us most significantly in the near future is the deeper implementation of edge AI.
So more data collection and more processing at the edge, which might include things like IoT devices, telemetry devices, but also includes decision making at the end.
And so, you know, I'm sure most people have kind of heard about all the kind of recent developments for even the implementation of small models on mobile phones.
And so you might start to see instead of having to move lots of data to more centralized points, you're going to start seeing models working where the data lives, or at least where the interaction lives.
And while there's going to be some opportunity in that, for better decision making and maybe more real time processing of things, things like it's also going to create a lot more data.
So, you know, it's a bit of a blessing and a curse.
So it has propensity to allow more value creation, more value harvesting, but it also says the propensity to bury you in more problems that you already have having too much, you know, unstructured data that's kind of already unusable and hard to manage.
So I think it just really just really takes, I think, some kind of focus and dedication getting that the very core and meaning goes.
So the core business value and use case for data and being very focused on that and maybe less about the technology and the tools that are around it because there's many good technologies to help solve the technology problems.
So just a super focused on the use case, the business value, how you're going to implement it, how you're going to implement the change it goes along with it, and how you're going to measure it.
I mean, I think if you get that recipe right, then some of the rest of the stuff is requires expertise, but it is somewhat secondary to just making sure you've got a really well designed system of extracting value from data.
We knew this could potentially happen before we even started recording because when you're talking about the edge AI and the small models, which are getting better and better and better.
I haven't played around with it yet, but even like the latest GPT 40 mini, the performance on the benchmarks and these, going back to Lama, the 7 billion and smaller models, how good they're getting right now that you can run on your local laptop and what that means for, you know, guests for big projects and, you know, general things using the frontier models is going
to be the way to go with their billions or trillion plus parameters and all that, but these small models are getting so much more cost efficient in thinking about data privacy and security and the idea of fine tuning those, but that's a whole other episode.
So yeah, it's well, I have to save that for a different day, but I guess, you know, and maybe we should have started before we just dove straight into the deep end.
I do think most listeners of this show certainly have gone out and they've done some work with these models, but I still have people asking me for prompt libraries, which just drives me crazy.
I think that's completely the wrong way to think about interacting with these models.
It's not, you know, if you want a prompt library, you should just have a big control panel where you're pushing buttons and say push the financial statement analysis button.
Instead, it's more like, I need to understand what this technology is and how to interact with it.
One thing that I encounter a lot is people trying to understand how to interact with these models and understanding the way these models were trained that RLHF, the reinforcement learning with human feedback, they're designed to be zero-shot prompt.
It's like, you give me a question.
I'm going to do everything I can to answer that whole question all at once.
And so if you default to asking a question no matter how big or small it is, it's going to try to answer it there.
But one of the things that you can see better results from is sort of the eat the elephant one bite at a time where instead of saying, do this massive thing, let's talk to it as we would an intern or another co-worker or an analyst we're working with.
But I know getting back, you know, out of the weeds of how these models work and how we implement them, just kind of a user to talk to me about the difference between zero-shot prompts and something like chain of thought prompting.
And any tips you have on how our users can use these interact directly with these LLMs to get the best results.
Sure. Yeah, yeah. So I mean, I think what most people are probably familiar with, certainly in their beginning experience with an experimentation, is zero-shot.
So you're basically given it an instruction.
Regardless of whether it's simple or complex or well thought out or not well thought out, it's really based on just a single step answer.
And like you said, it's going to give you everything and its power to be able to answer that.
Maybe if you've given it some conditions.
So it kind of goes wide. And based on some settings you can change in the LLM, it might go to wide, it might go to narrow, it's, you know, kind of less predictable.
And then if you asked it the same question again immediately after you cleared your cash on the first one, you might get a different answer.
And so there's got a lot more variability, but I think it's a great place to start.
And then what I would say that where the chain of thought prompting is more significant is as you're refining.
So think of it more of a, as a conversation that you're having, you've pre-sequenced it.
So you're not actually doing it real-time, but think about it as steps in a conversation where each of the outputs of the prompt, because they are still in the chain zero shot, right?
But each output from the results of the prompt are being used to refine and improve the output of the next sequence in the chain.
And based on the way LLM's work with memory that's involved, it starts to construct a more refined understanding, because the data that's being used to prompt the second, that's being used in the second prompt in a sequence is now informed by the results of the first one and so on and so forth.
So it's technically getting narrower as you, as you make your progress through the steps.
And so you might think from a, like from an FPNA standpoint, you say, well listen, let's speak very colloquially, but chat GPT or GPT, tell me, you know, what the forecast is for, for next quarter.
And then, okay, now we have that result.
You say, well, now let's add in the previous three quarters.
So I want to start looking at comparison.
And then tell me what from the original forecast to the actual results of Q1, Q2, and Q3, where were the variances?
And now, for now, now let's go to the next sequences.
Now that I know what those variances are, let's add in all the cost predictions to the fourth quarter.
And then the next sequence you're saying, now take those same things that were responsible for variances in Q1, Q2, and Q3 and apply them, knowing these cost impacts that are the forecasted for Q4.
And tell me if you see any, any, any correlation or any variance, anything I should be aware of.
And then at the very end of the final sequence, you say, now take all of this information I just asked you for and write a Q4 forecast prediction summary for me.
So as you can see, you're kind of having that conversation.
And like you said, it's almost like you're asking, you're giving directions to someone who's on staff or even yourself was saying, like, here's the things I need to collect.
And then, but it's actually providing the analysis automatically as it's going through it.
And so, I think, hopefully, I give a good example of the difference of uses.
I think people initially were kind of instructed that you should spend a ton of time perfecting the zero shot approach.
And I have seen 2000 line prompts, like amazing stuff.
And maybe that does work.
So I'm not really kind of commenting on that.
But I think what I've had better luck with in creating the the chain of thought approach is actually reverse engineering.
So if I tell the LLM what I'm trying to accomplish and how I want the end result to be prepared or demonstrated, I can start asking questions about what it needs to help me fulfill that text.
And if you start going kind of like a reengineering recursiveness through the sequence, you can get to the origin.
And then you can take that sequence, it mapped out and create your chain of thought prompt.
And so it's at least a little bit different way of arriving at the same result.
So either if you kind of really already know how you would define the sequence, have at it and set it up as a chain of thought prompt.
But if you're not really clear on how to get certain bits of the information and the sequence correct, it will hope you sorted out as long as you can tell what your goal is.
It will help you start going backwards through the sequence of getting to the origin, which can be then the origin of your chain of thought sequence.
Yeah. And it's real, if you think about it, it's the way we would solve a problem.
If somebody gave me, you know, when I made it do that, I would go through the same steps that you described for doing the forecast and all of a sudden I need to understand this and do it.
So, you know, because there's also different approaches, I guess if you went with that zero shot, if you wanted to come up with a new way to do it, rather than going through, well, this is the way I would do it.
That might be a its own use case is say, well, what would you do with the forecast?
And then, you know, feeding it different information.
But the danger of zero shot, I guess, is a lot of times that you don't give enough information up front for it to drive those decisions.
Yeah, agreed. And that and just, it is more generalized.
You're never going to get to quite as precise and output.
I think if you're if you're diligent about the chain to thought sequence and prompting, it's it's just really, it's just two different use cases if you ask me.
Yeah. All right. In the interest of time here, I'm going to we're going to make a hard right turn.
I got two more questions for you.
So, first off, and I always love this one because you never know what you're going to get.
You get answers all over the board.
But the first question is what's something that not many people know about you?
Probably because I spend a great deal of my time both personally and professionally talking about technology is that I am a very big fan of the outdoors and and the awe that comes associated with seeing really beautiful landscapes and outdoor environments and have a very deep appreciation for actually turning technology off.
Yeah, you got to have that balance, right?
Yeah. And you're in you're in Arizona, right?
So you have plenty of scenic beauty all around you.
So that's great. All right.
Now this might this one's going to seem weird.
I think considering probably where you spend most of your time.
But again, I ask I ask every guest this and it's it's always interesting to hear what people come back with and asking you as someone who probably doesn't spend a whole lot of time in Excel.
What is your favorite Excel function and why?
Yeah. So while I spend plenty of time in Excel or throughout my career and and of a deep love and appreciation for pivot tables, you know, my answer is actually going to be a simple one and it's just using the filter tools.
So for me, when I look at arguably unrelated to kind of the data science side of things, I'm a big drill down guy.
I'm looking for, you know, data matches.
So if something occurred at a higher level, if I keep drilling down, do I see the same pattern as an example or do I see it across, you know, kind of let's say different, you know, different instances of different funnels and in what is ultimately in the Excel tables itself.
So I'm just a big nerd of changing the filters on things to continue helping to answer those questions.
So if I saw a phenomenon in this business at the top level, was it across all geographies?
So I can filter that. I can start looking and then if I find two or three geographies where I did see the same thing, now I can filter it again and look for was it, you know, customer-related or product-related.
And so I'm a big drill down guy and so kind of live and die by the ability to toggle on or off all sorts of different filters.
I love it. And as a stats guy, you know, I'm guilty of this, but it's like p-hacking where you're going through and you're like, well, this is the data I have.
Let me, you know, you're not, it's not what you're testing for, but you find things.
So, but because we're not in academia, and we are just in doing business data analysis, I have actually found value in doing that very thing where it's kind of like finding those correlations where you just drill down and drill down and say, you know, it does the pattern repeat or is there a new pattern that I didn't see before?
So, yeah, I'm right there with you.
So, all right. Well, Brandon, I really appreciate you coming on the show.
I guess finally, how can our listeners connect with you and learn more about you and your work?
Yeah, sure. So my company is Steady Dynamic.
We're based in Phoenix, Arizona.
We've got offices in Warsaw, Poland as well.
So we're a global company.
Our website is www.steadynamic1word.com.
And if people are interested, you could always find me on LinkedIn under Steady Dynamic, Brandon Wilson.
I think my LinkedIn is BK Wilson, but you'll find me if you visit Steady Dynamic and be happy to connect.
And also, Glenn, thanks for having me.
I really appreciate this.
I know, again, we probably could have gone much wider.
So we'll have to organize the second one, but this was really enjoyable.
And I hope the listeners benefit from the conversation.
All right. Thanks so much, Brandon.
All right. Have a great day.
Thank you.
