Artificial intelligence (AI) is quickly becoming an intricate part of the decision-making process within financial firms. Initially used for simple tasks like summarization and drafting reports, Prasad Shinde, pre-sales senior engineer at KX, tells Trader TV that AI is being used to streamline the entire ideation-to-execution pipeline, whether it’s for trading workflows or across the entire business.
Delving into the value of temporal AI, Shinde shares his views on why it’s important to know when to use data as much as the actual data itself in real-time trading environments.
But applying temporal AI isn’t always straightforward. The senior engineer unpacks both the hardware and software-integration challenges that comes with these AI-powered solutions.
Looking ahead, Shinde shares his predictions on where he see AI agents with natural language interfaces making the biggest impact to financial firms businesses in the future.
Transcript
Jo Gallagher Are you seeing a growing adoption of AI solutions across financial services, and if so, what’s behind that momentum?
Prasad Shinde The clients want to gain a competitive edge in the market, and to do this, they have to be faster than their competition. AI is delivering on the same mandate for our clients. What we see is clients moving from low-hanging fruits, such as summarizing, drafting of emails, and giving efficiency plays to their employees by using coding assistance or virtual assistance. You are now seeing clients using from ideation to hypothesis squishing that time taken to deliver on an idea. For example, if you saw virtual assistants, clients are using Erica, which is a virtual assistant which has had 2.5 million (billion*) client interactions and sort of many billion touch points from various clients. That’s Bank of America’s virtual assistant with a sort of making driving towards revenue generation part of it. JP Morgan’s Index GPT is very widely talked about, which has helped you to build sort of a thematic basket of stocks based on the themes and keywords gleaned out of those using various models. That’s where I see the adoption and that’s why KX is helping a lot of our clients. [Note: BofA reported that Erica surpassed 2 billion interactions back in April 2024)
Jo Gallagher Really interesting. And so why is when you use the data just as important as the actual data itself in AI-driven trading decisions? And how can this temporal dimension also create some competitive advantage?
Prasad Shinde Data as it generates, loses value the more you delay gleaning insights from it. And this is what I’ve seen in interactions with our clients. A lot of them historically seen that AI has static nature of data for their training and validations in the current climate where the markets are volatile. By the time you put your model into sort of execution mode, your data it is trained on is already stale. So your data or your AI model needs to be a stream almost of real-time what’s happening in the economy or the market. Typically, you need to know the context of why it has happened, how fast, or what are the peculiarities of moving towards understanding causal nature of things rather than just correlations. That’s why our clients are aware of the temporal nature of the data, and that’s why they’re making it available for their AI models so that they can glean insights faster.
Jo Gallagher What are some of the challenges now of actually deploying these temporal AI-type solutions, particularly in real-time trading environments?
P Yes, there are a lot of challenges. Some of them are hardware-related, some of them software software-related. I’ll talk about the software-related first. Traditional databases typically would store the data and it is not very efficient to kind of like take the data for analytics in terms of you pull the data into memory. If you’re looking at a lot of traditional databases, you’re going to have to put a lot data into the memory. It becomes expensive, it becomes slower. You have to keep on frequently adding capacity. This is where KX is helping clients around being able to provide a high-performance computing platform where you bring in exactly relevant data into the memory for your calculations, aggregations for your model to kind of like use that and leverage for making insights or inference cycles. The second part is around your CPU (Central Processing Unit), GPU (Graphic Processing Unit) sort of like interplays historically would see lot of architecture on CPU. Which with the advent of GPUs, now you can see the faster rendering of your models and your sort of like hypothesis and stuff like that. There’s still some gaps around the interplay between the rendering instructions for GPUs are provided by CPUs. So that interplay between CPU and GPUs needs to be faster, which is what our partners such as Nvidia is trying to do. And it has come up with new versions of chips such as Blackwell, which kind of like make it more seamless, more integrated to kind of have that ecosystem. For a faster processing of data across for you to innovate, ideate, and sort of execute the ideas faster.
Jo Gallagher Where do you see these types of AI solutions making the greatest impact in financial services in the short and long term?
Prasad Shinde That’s very topical currently. In the short term, I see the whole point where a quant analyst is trying to ideate something. For instance, if you take tariffs, it might be causing inflation in the North America side. If that is your hypothesis, you might want to hold value stocks. If this is your hypothesis you want to kind of like bring in data which is historic in similar times, also compare it to something which is real. Build a model or testing this hypothesis of yours, Provide that data in real time or as fast as possible to train it, have a separate data set for validating this unit to improve accuracy, and then put it on data in real time to infer whether your model works in real life. Historically, when I was an analyst at Bank of America, it would take me ages to get these data sets to kind of like piece together something of a trade idea, which is probably executable. Now that stack is coming, not just in days, to sort of like hours where a quant can go in, pull the data out, submit it to your model, infer from the model and get a very impressive cost-to-performance sort of inference and execute on that idea. This is the short term. In the longer term, you will see agents driving a lot of this with a natural language interface where a business user like you and me can kind of put in an idea. It will glean which, say for instance, symbol it is talking about, where that data is stored, call another agent which can take the data out. If there is a calculation involved, while pulling the data out, the original agent can bring in other sort of calculation, calculating agents. Combined there’ll be a aggregating agent which can pull together, piece together this information for all these, give you the desired results and you can act on it. So again, for the squishing of the stacks from ideation to execution. That’s where I see the world moving towards.
Jo Gallagher It was a pleasure speaking to you Prasad.
Prasad Shinde Thank you, Jo.
Jo Gallagher I’d like to thank Prasad for his insight and of course you for watching. To catch your other shows including Trader TV This Week, every Monday at 6.45 UK time, go to Tradertv.net
Catch Trader TV shows each week:
To watch Trader TV This Week shows go to – Video
Or make sure to follow us on LinkedIn – Trader TV’s LinkedIn
Available as a podcast – here
