Hi everyone. I'm Hil Kati, a software engineer focused on AI cloud and FinTech. Today I'll be walking you through how AI powered digital twin technology can transform the way we manage financial risk and optimize assets across industries, especially in high stake environments like finance and operations. Let's begin by understanding the context behind the stock. In today's fast evolving financial landscape, organizations are under constant pressure to manage risk, performance, and efficiency all in real time. So traditional approaches that rely on static models, periodic reports, or siloed data are increasingly insufficient day by day. So this is where our digital twin technology becomes a game changer for us in the financial world. A digital twin is a real. Time. Digital replica of financial entity or a system. It could represent a portfolio, a payment network, risk model, or even an end-to-end business process. Unlike static dashboards, these twins are. Dynamic con, continuously updated and also simulate both current and potential feature. Also like the states using live I would say like live data streams here. So now when we integrate this with artificial intelligence and machine learning, we unlock the ability to predict. Financial anomalies and also risks before they materialize and also simulate complex market and also credit scenarios. And another thing is to continuously optimize asset management and decision making at scale. And now I would like to walk you through a framework that combines digital twins, real time financial data, and also a IP work pre predictive models, including. Random Forest classifier. A powerful I would say like it's more I was saying like powerful because it's more based on a learning method for classifying risk patterns and financial events. And another one is L Esteem, which stands for long short-term memory networks, which is like a type of recurrent neural network. Particularly effective for detecting patterns over time in financial time series kind of data. So this fusion of technologies empowers financial institutions to detect irregularities early and also allocate resources more efficiently and also move faster and make more informed decisions on a timely basis. So in essence, this is how we move from. Data to decisions using AI and digital twins to minimize risk and also maximize the value. And coming to the background, what we are trying to do here is it's like a more of a concept evolved and why it matters today in regular finance is the most important. So now getting there, digital twins began in aerospace and manufacturing, I would say as digital replicas of physical assets that initially helped. Simulate and monitor, improve their behavior. But over time, the concept has evolved beyond the physical world. Today we apply digital twin technology to finance to create virtual real time models of portfolios, customer segments, credit flows, and operation pipelines. These twins. As living financial models continuously update using transactional behavior and also the market data. So now when we pair this artificial intelligence and machine learning model, we unlock next level capabilities. That is more of like forecasting or risk exposure before even they cause harm. And another thing is to detecting deviations in real time. Like it can be like fraud or even the default risk even that can happen within a twin of. Your actual financial system that you have created. So this shift reflects a broader trend, part of Industry 4.0 and FinTech 2.0, where traditional systems are giving away to automated, adaptive and intelligent infrastructures here. So I'd say instead of reacting after risks or inefficient appear, organizations can use digital twins to anticipate and also act ahead of time. So now there, obviously for every thing that we work on, there's always like problems that we come across. So looking at the problem statement and objectives that we would like to identify as part of this talk are presented in the next slide. So before we explore the full architecture and the process that's being done as part of this overall digital twin technology, let's define the core problem that we are trying to solve here. So most financial systems today still rely on after the fact analysis, whether it's like risk scoring, asset performance tracking, or operational reporting as well. So these systems often. Operate in silos like there's always like a blind spot and also delayed reactions in critical moments. So even where digital twin models exist, they're usually static fragmented, and also like it can be domain specific most of the time. The most often things that they lack here are like real time adaptability and also like seamless integration with live financial data streams, and also the intelligence to learn, evolve, and optimize themselves over time. And also when these financial activities are interconnected. So they form credit risk to liquidity management and also asset performance of regulatory complaints. So to address this, our objective is to build an AI power digital twin framework that brings real time awareness across financial systems and also a driven insights through models like RFC and LSTM that I have discussed earlier. And the ability to simulate and respond to scenarios proactively, whether it's fraud detection, market shifts, or like asset financing, everything can be more likely processed with some good data collection and how we pre-process the data. So that is something we can look into the next slide. So coming to the data collection and pre-processing, let me walk you through how the data processing or overall collection can help data with our digital twin models. I would say more of helping a real time financial data streams. So we begin by tapping into real time financial data streams, which include transactional data from payments, loans, and trades. Portfolio level metrics, like if you say the portfolio level metrics, there are a lot more risk involved in that, like exposure risk scores and also a lot of fluctuations. So coming to the indicators where we can identify this, like there are always like liquidity indicators, both internal which I can say like reserve ratios and external like funding availability. These are like some of the. Financial data that we need to collect as well. And coming to the market data such as volatility, interest rates, or economic sentiment indexes pulled from a lot of different APIs are also collected and which is very critical for us. So as part of the data cleaning process or like pre-processing, the first step is to conduct data cleaning. Handle missing records and also detect anomalies like spike in transaction volume. That may be a system error or like it can be also resolving or formatting mismatches in general, like it can be like a lot more anomalies too. So then we perform normalization and feature scaling so that all inputs are aligned in a compatible range. Which is crucial for most machine learning models like random forests or L SDMs to learn accurately and efficiently. So coming to the proposed framework. This is like a proposed framework, which are like step by step done in a defined way to, to get the overall data cleaning and also the pre-processing. As we see, this slide shows that end-to-end architecture of the AI power digital twin framework tailored for. Financial risk management and a sub optimization. Here we see how the data collection happens. So the data collection begins by collecting real-time financial data. So we, that happened from core systems, like as mentioned, like payments, loans, credit platforms, and also the market feeds, interest rates, volatility indexes, and also. APIs and third party providers, which can provide complaints, updates, and also the sentiment signals. So this serve as the foundation for understanding life, financial behavior, and risk exposures or outliers that we can always look for. After this, we do the data cleaning to correct or exclude all the flawed records. So this is where the cleanup happens. And now, and the next step is to normalize to this like align different scales, crucial for training relatable models that we were discussing in the previous slides. So the next step is the feature engineering. This is where the extract of predictive indicators happen. So rolling averages or transactional patterns or like liquidity shifts over time. Anomalies in portfolio drift or credit realization and also like synthetic risk flags based on loan three sh also happen in the future engineering step. And then we do the digital twin model development here, a real time enhanced digital replica of a financial system that has like. Portfolio manager, credit underwriting flow or fraud detection pipeline. These kind of pipelines are built here. So this twin evolves continuously as new data arrives and it gets integrated with the financial telemetry instead of the physical systems and also simulations of cash flow and also like market turbulence can also be monitored and. As said, these ML models are trained on historical plus streaming data. So now when we see this random forest are used for rule-based risk classification and like LS, TM is used for time series predictions, example market response which happens over the time. And also like we use SVM model for separating risk tires at different levels, and we use auto encoders for anomaly detection without label data. So as we see the asset management happens. More like into this insights feed directly into capital allocation, liquidity balancing, or fraud handling strategies in a more dynamic manner. So when we come to the reinforcement learning here, so finally, I. When we are applying the reinforcement learning to continually improve the system, it learns from our outcomes to adjust thresholds, reallocate capital, or also optimize strategies based on the rewards feedback that like cost savings or risk mitigation success, anything that gets back to the system. So in short. I would say like this framework connects data to models and models, get connected to the financial intelligence, and which in turn measures a lot of business value in terms of different aspects as part of overall process. How does methodology and overall this framework worked? So there, there are like some results that we got popped up once. We are, once we are processed through that system, so now when we look at the results, how it performs, like looking at the bar chart, we are comparing four machine learning models here. The random forest classifier, long short term memory and support vector machine. And also like the auto encoder. So each was tested on a financial data set to predict risk signals such as anomalies in transaction flow, portfolio stress, and also credit utilization drifts that keeps happening. So coming to the accuracy here, so RFC is leading with 95. 0.5% followed close by lstm, which is a bit shorter than that, which is vital for realtime decisioning here. And coming to the precision RFC and LSTM models, again take the lead indicating their ability to reduce the false positives. So this is very critical in finance systems to reduce the false positives, like I would say, like for example, flagging fraud or credit risk only when it truly matters is really important. So coming to the recall part of it, so again here, RFC and LSTM also excel with a lot more higher percentage compared to the respective other models. So here, high recall means fewer false negatives even, which is that which is also more important in financial risk analysis, which is important also for catching all genuine risk signals to understand if it is really a false positive or a false negative in coming to the F1 score. Fund score is also indication for a solid balance between precision and recall. So if you see this even here, RFC in L-S-D-M-R at a higher level indicating a really good balance, as I was saying. So auto and coder and SVM performed reasonably well, but are a be ru variable in balanced trade offs when we are trying to make some trade offs between different models. So coming to the. Overall results, the precision recall and accuracy that is all important, that when we are looking at the pre-call recall precision recall and accuracy, let's now break down like how the performance of each model across four evaluation metrics, precision recall, fund score, and accuracy are monitored. So these metrics. We're calculated based on each model's ability to predict financial anomalies, like unusual transaction patterns, risk threshold breaches, or also early signs of asset under performance. So when I say like random forest classifier, so this was achieved by highest across. All categories, accuracy, precision, recall, and also F1 score. When we are looking at the balance between these two, it helps detecting true risk events and minimizing the false positives here. So long short term memory, LSTM model, like each performs very close to RFC and also like especially strong in recall making it effective for capturing. Time-based financial patterns such as liquidity depths, and also like credit defaults over time when we are coming to the next model support vector machine. So slow, moderate performance accuracy is like 90% lower recall, such as it's likely to mess risk signals over time. So still valuable. For binary classifications like fraud detection and also with clear patterns. So when it comes to the auto encoder, we, it also performs well considering it's unsupervised with the more precision value, but also compared to the anomal as it detects. So when we are coming to the final decision that we want to make both RFC and LSTM. Stand out best performing models for real time financial monitoring and also like risk prediction offering the accuracy and balance required for high stakes decision making in finance. So looking at these models, there are lot. There's another thing that we need to focus on discussing like overall like cost benefit coming to how these models performed. So when we're trying to discuss the cost benefit. So I'd say like the initial cost would be like a little initial cost. To make this overall digital twin perform better would be like little lesser. So ideally, like it would help, like overall downtime reduction, like traditional setups can't adopt in real time, so they offer 0% downtime reduction. But in DT enabled models. We improve system resilience and prediction accuracy leading to a 25% reduction in downtime, like I would say, especially during high risk events like liquidate shortages or fraud spikes, and however, the return on investment. So traditional systems ideally don't yield. I. A meaningful return on investment in dynamic financial environment. But the DT Powered Systems by Contracts show a one 45% return on investment, validating not only their operational value, but also their financial justification for their I would say like more of a decision making situation. Coming to the final touch point, what I wanted to say is digital twins require great investment upfront, but they generate substantial recurring value over time, both in financial returns and also operational silence. So to wrap up, the digital twin framework clearly demonstrates how AI integrated system with digital twins can elevate financial systems by making them more predictive response and also silent by analyzing the approach we have achieved. I would say like very much measurable benefits, a 25% reduction in unplanned operational downtime, an 18% decrease in overall maintenance or overhead costs, and also like a 20% improvement in asset utilization where. That helps with the capital efficiency, liquidity deployment, or also risk-based resource allocation. This is possible just that we are not just monitoring system anymore. We are modeling them intelligently in real time. So when I say the random forest and long short term memory models, we are able to generate high confidence predictions and simulations. These predictions directly support strategic decisions from mitigation. Also to portfolio rebalancing or also liquidity allocation. So as a result, I, we have seen like one 45% return on investments, which proves that it isn't just a technical win, but also it's a financial one. And also these intelligent systems also enable real time visibility and automated scenario planning, which helps financial institutions take action a bit early, avoiding crisis, and also can definitely improve the performance. However, as the challenges remain, the scalability is still across large, multi-layered financial systems, right? And the real time synchronization, especially when integrating diverse. Data sources. It gets a little difficult, but it's still manageable. System compatibility is also how well the system is compatible with the legacy infrastructure. That is one thing to really consider here as these are more of like a solvable problems and they don't really point to the need for continued innovation in cross-functional collaboration to unlock the full potential of AI powered digital twins in finance. The future of financial Resa lies in more of moving. From static reporting to living learning systems and here the digital twins powered by AI are really the bridge that can get us there. Thank you so much for the attention. I'm all good and open to answer any questions that you have here today.