For over two decades, Gaurav Gupta, Managing Partner at Decimal Point Analytics, has watched capital markets evolve from paper ledgers to spreadsheets to agentic AI systems that now read balance sheets, extract data from annual reports, and flag risk in real time. He sits at the intersection of institutional finance and applied AI, building customised models for clients managing assets worth hundreds of billions of dollars.
In this conversation with CIO&Leader, Gupta unpacks the architecture behind Rakshak, DPA’s behavioural AI system designed to nudge retail investors away from emotionally driven decisions, and explains how the firm navigates data residency across the UK, US, and India without compromising model performance. He also addresses the hallucination question head-on, describes where AI genuinely generates alpha beyond human capability, and offers a candid assessment of how close — or far — the industry really is from fully autonomous portfolio management.
Managing Partner
Decimal Point Analytics
CIO&Leader: Rakshak uses behavioural AI to nudge retail investors in real time — what’s the underlying model architecture, and how does it detect emotional decision-making at the point of transaction?
Gaurav Gupta: For customers, what really goes on at the back end is a wealth of information and data about the investor. On the other side, we’ve built fairly extensive sources of input for various investments and stocks, with proper categorisation and classification of their specific nuances—whether it’s the industry, the product, the supply chain, and so on. All of that is properly identified through the system.
On top of that, we continuously monitor day-to-day news that gets linked to the relevant stock. Based on this, a proper risk level is assigned to each of these investments.
CIO&Leader: How do you train behavioural AI on financial data without introducing the same biases that cause poor investor decisions in the first place?
Gaurav Gupta: The behavioural data is closely linked to how we classify investors and identify trends in their investment patterns, income levels, and related factors. So, the behavioural aspect is more focused on classifying specific categories of investors and overlaying that with what could be suitable for those classes of investors on an ongoing basis.
CIO&Leader: DPA operates across the UK, US, and India — how do you build AI models that stay compliant across three fundamentally different regulatory regimes simultaneously?
Gaurav Gupta: Actually, Decimal Point Analytics has been in operation for over 20 years. Our interactions and customer base are predominantly institutional, and many of them are fairly large-scale. Our largest client has close to half a trillion dollars of assets under management, and some are even larger, although we work with their subsidiaries.
For this class of customers, we have to build very customised models to address very specific problem statements. Every model is built to address specific operational areas and workflows.
For example, we have a client that conducts extensive lending across the United States for small business owners. The task involves processing large volumes of documents, including tax filing papers and balance sheets. These are small companies, not large organisations with structured financial databases.
We have built systems to extract these datasets, much like what we’re doing in Rakshak, where the effort is to extract data from annual reports and populate a structured template. In this case, we’ve applied local knowledge because all the models must be fine-tuned to process those documents, extract data, perform spreadsheet analysis—what they call “spreading”—and conduct credit analysis using our agentic models, which have been built specifically for certain categories of borrowers.
These are highly customised models built and fine-tuned to meet specific customer requirements.
CIO&Leader: Private equity and hedge funds live on proprietary data advantages — where exactly does AI generate alpha that human analysts structurally cannot?
Gaurav Gupta: It’s amazing the kind of work that’s being done. Of course, the largest and most sophisticated fund managers have their own secret sauce. But from what we’re doing, it’s already very interesting.
One example relates to what I mentioned while describing Rakshak and matching investors’ risk-bearing capabilities. We use a similar approach in another application where investors want to understand what kinds of investments their competitors have made.
For example, if competitor X was successful with a particular strategy, they want to understand what triggered that investment two, three, or four years ago, and what made them choose that sector or stock that eventually generated returns years later.
The amount of data that needs to be processed, organised, and categorised is enormous. That simply cannot be done through a normal analytical approach.
This gives insights that I don’t think are possible through human analysis alone or even with Excel spreadsheets. You need massive datasets, significant processing capability, and the ability to relate different datasets to one another. You also need models that can identify signals very early and detect a particular direction or investment opportunity.
These are the kinds of analyses that humans simply cannot perform through traditional spreadsheet analysis.
That’s why many investment managers today are using AI to surface the most impactful information, allowing them to quickly identify potential trends.
I don’t see this as a case where AI wins, and human analysts lose. That’s not how it works. AI systems augment the work of fund managers by helping them generate better insights for their portfolios and make better-informed investment decisions going forward.
CIO&Leader: What does DPA’s data pipeline look like for alternative assets, where structured data is scarce, and valuation is inherently subjective?
Gaurav Gupta: In those cases, let’s take securities or assets where valuations are inherently subjective. For example, evaluating private equity companies often involves extensive forward-looking projections. We’re now seeing this even in secondary markets around IPOs of companies like Tesla and OpenAI, where projections extend so far into the future that they almost resemble venture capital investing.
The objective isn’t to say, “This is the right valuation,” especially in such subjective environments. Instead, the role of data and models is to highlight the extent of risk being taken at a particular valuation level. In other words, what is your risk capital on a particular investment at any given point in time?
These models ingest extensive datasets, including macroeconomic indicators, comparative company information, analyst opinions, and market sentiment around a potential investment. All of that is brought together into what you could call a model driver.
Ultimately, however, everything has to be evaluated in terms of risk—what level of risk is being assumed and how far the valuation extends beyond what is fundamentally supported at that point in time.
CIO&Leader: LLMs are now being used for financial research — how do you solve for hallucination risk when the output is an investment decision?
Gaurav Gupta: I think the models have become extremely good. Concerns around hallucinations were much more relevant about 12 months ago, but nowadays the models that have come out are so finely tuned for specific purposes that hallucinations are quite limited in that respect.
As I mentioned earlier, we fine-tune our models for specific kinds of analysis—for example, credit analysis for certain categories of small companies. The ability to fine-tune different models is one part of it. Beyond that, there are different ways of creating guardrails to ensure the models don’t go off track. Neural networks can sometimes behave unpredictably across different scenarios, but the ability to create those guardrails keeps things within a manageable range.
In that respect, these models are becoming increasingly effective at augmenting analysts. At the end of the day, there’s always a human in the loop. There’s always somebody overseeing the models before the output is released as the final outcome, at least in many use cases. If you’re talking purely in terms of equity, certain systems can be fully automated. But in areas like equity analysis, where a lot of interpretation is required regarding various environmental factors, the human is very much in the loop.
CIO&Leader: Rakshak delivers real-time behavioural nudges — what’s the latency requirement, and how do you architect for that at scale across millions of retail transactions?
Gaurav Gupta: This is still an idea that’s being built, but we haven’t rolled it out to many investors yet. However, in today’s environment, we can create highly effective, fine-tuned models for specific purposes.
These are very efficient models. If we build them correctly, they can be converted into very effective edge-based systems, which means they don’t require heavy compute. These systems are designed so they don’t need to keep calculating endlessly at multiple levels, which is also an important guardrail.
So, we’ve created this system to run without requiring extensive internet connectivity or heavy GPU-based infrastructure. That’s the approach we’re taking with this particular model.
CIO&Leader: How close is the industry to fully autonomous AI-driven portfolio management, and what’s the last technical barrier before human fund managers become optional?
Gaurav Gupta: In many respects, human fund managers have already become optional because we have a very large ETF industry that sometimes delivers even better performance than human portfolio managers. However, that generally happens under specific market conditions—when interest rates are low, and the markets are broadly bullish.
When there are economic headwinds, geopolitical uncertainty, or the kind of volatility we are experiencing today, the models are still not ready to support rapid decision-making at an acceptable level of risk. The Sharpe ratios would simply become too volatile, and the overall risk would be too high.
So the models are not yet ready to completely replace human portfolio managers. I wouldn’t want to predict how many years away we are from that situation. When we speak with our customers, we always recommend keeping a human in the loop, especially for critical decision-making systems.
I think we are still some time away from being able to completely replace human input and oversight.
CIO&Leader: Cross-border AI in financial services means cross-border data flows — how does DPA handle data residency without fragmenting model performance?
Gaurav Gupta: As I mentioned earlier, most of our customers—whether in the United States, the Middle East, or Europe—are extremely sensitive about their data. Forget cross-border movement; many of them don’t even want their data leaving their own systems.
So, we deploy most of our models within the client’s environment. It is cloud-based, and we access it through secure VPNs, but the models remain inside their infrastructure.
Today, everything is in the cloud, and we deploy our solutions accordingly. To comply with regulatory requirements, we ensure that the data remains within the required geographic boundaries. The cloud infrastructure, networking, and processing capabilities are mature enough that it doesn’t create any operational challenges for us, whether we’re working from India or any other location.
At the moment, this is not a significant challenge for us.
CIO&Leader: Two decades in capital markets — what’s the one financial workflow that AI has genuinely disrupted, versus where the disruption narrative is still mostly hype?
Gaurav Gupta: I think AI has definitely eliminated the need for large numbers of young analysts to manually go through documents and extract data. That process has become highly sophisticated.
If you go back three or four decades, when I joined the industry, computer hard drives were roughly about four MB. Just to put that into perspective, today we have machines with 8 GB, 16 GB, or much more RAM, and storage capacities are exponentially larger.
Back then, people used to write their financial analysis on paper. That was eventually replaced by spreadsheets, which brought in a different generation of analysts.
What AI has done is significantly reduce the amount of time and effort required to process information at the first level. It has fundamentally disrupted that workflow. Automated spreadsheets are now being created, while the final analysis is still done by humans. However, much of the preparation is completed well in advance, particularly in equity markets and equity investing, which seems to be the focus of our discussion today.
That part has already been significantly disrupted, and it will continue to evolve.
Earlier, most of the available data was structured and could easily be fed into Excel through different data connections. Today, we’ve moved from structured to unstructured data. News articles, economic announcements, and other forms of unstructured information are now being processed by AI and brought together into a unified analytical framework.
That entire process has been substantially taken over by AI. I would actually be apprehensive about professionals who are not using these capabilities adequately because they will increasingly be left behind.
This version is publication-ready while retaining Gaurav Gupta’s conversational tone, technical terminology, and intended meaning.