Hi my name is AK K with overall 18 years of IT experience specialized in data engineering background. And I most of the time work with data. I would like to thank comp 42 Machine Learning 2025. But giving me this opportunity to speak in this conference. So without much delay, let's dive into AI unhinged fraud detection in financial services. This presentation explores the implementation of an AI driven fraud detection system at a leading financial institution. The solution integrates diverse data sources. Applies sophisticated machine learning algorithms and also enables real-time transaction analysis through a comprehensive cloud-based architecture. This transformation from traditional rule-based detection to a adaptive multi-layered framework, significantly improved fraud prevention while also enhancing customer experience. Let's talk about the evolution of financial fraud and also the challenges that are being faced. During evolution there are four main challenges. The first being high false fossil rates, the second one being siloed data systems. Third one being significant detection delays. And the last one being customer experience impact. These inefficiencies resulted in annual fraud losses exceeding 25 million, and also deteriorating customer satisfaction metrics. According to research organizations with siloed data systems, miss critical connection patterns that could identify 27.4% of complex fraud scenarios, Industry fraud landscape. So there were around four areas where you could see in this slide there are four various areas where we have focused for gathering some of the metrics financial services, retail and e-commerce, healthcare and government. According to the Association of Certified Fraud Examiners is EFE two in 2022 report to the nations organizations. Lose an estimated 5% of revenue to fraud each year, which translates to global losses of approximately 4.7 trillion. Financial services remain the most targeted sector with the median loss per case, reaching 1.5 million and cases lasting an average of 12 months before detection. So you could see some of the metrics that have been highlighted across various areas. Like annual fraud losses the median loss per case, and what is the average detection time, the areas being financial services, retail and e-commerce, healthcare and government, Technical architecture, cloud page, realtime analytics. If you see here in this slide. Around 1 67 0.3 milliseconds. Average crossing time has been done for the real time station engine. And there were around 237 behavioral indicators analyzed in machine learning pipeline. And finally in the data integration framework, there were around 17 disparate systems that were unified. The solution leveraged cloud infrastructure with 99.99% availability. S. Elastic computing resources that could scale from 5,230 to 25,004 70 transactions per second during peak periods. The A CFE report notes that organizations implementing cloud-based analytics solutions with real time monitoring. Different fraud schemes and an average of 33.2% faster than those using on-premise legacy systems, resulting in 47.8 lower media losses. The data integration framework, there were three main layers in this framework. The first one is data standardization layer. The second one is Apache Kafka Kafka Streaming, and the last one being Data Lake Architecture. The data integration framework unified disparate systems including core banking, credit card processing, digital banking channels, and customer relationship management. This consolidation eliminated data silos that previously concealed critical fraud indicators. Within the data standardization layer we processed around 7,000 8 34 transactions per second with ETL pipelines normalizing 14 different data formats using 2 83 transformation rules and maintaining 99.99 percentage accuracy With the Apache Capish streaming we deployed across 32 broker nodes with 42.7 millisecond average latency. Managing 47 topic partitions with 3.8 GB per second throughput and 99.9% delivery guarantee. Within the data lake architecture we ingested close to 17.3 terabytes daily while maintaining query response times under 1.2 seconds for 98.7 percentage of analytical workloads with 37 months of historical data. The next slide talks about machine learning pipeline. There were four core areas the first one being feature engineering. The second one is a continuous retraining. A third one we have ensemble modeling. And last one being the annually detection. So the ML pipeline represented the analytical core of the fraud detection system, leveraging advanced techniques to identify suspicious patterns across massive transaction volumes. The ensemble modeling approach combined supervised learning algorithms with unsupervised techniques for comprehensive detection capabilities. There were around 2 37 behavioral indicators across transaction characteristics. User behaviors, device attributes, and network patterns within the continuous retraining. We had got a feedback loops from 270 3004 29, confirmed fraud cases monthly with updates every eight hours in the ensemble model, in the ensemble modeling around six supervised learning algorithms with many times. Boost showing superior performance around 94.7% precision. And lastly, in the anomaly detection, the multi layer approach, identifying sat statistical outliers across 1 47 dimensions simultaneously, This slide highlights about the real time decision engine the real time decision engine. Sorry. Transformed analytical insights into actionable fraud prevention through a high performance architecture capable of millisecond level decisions. This performance level enabled realtime intervention before funds left the institution in 94.7% of fraudulent attempts compared to just 2.3 percentage with the previous systems, Implementation results and. The system demonstrated remarkable implements across key performance indicators. Customer satisfaction scores improved by 22.3 points on a hundred point scale with the percentage of customers reporting transaction friction declining from 28.32 7.1 percentage. Net promoter score recovered from 33 to 58, significantly exceeding the ministry average of 45. The, there were some of the numbers that you could see here some of the key performance index, like the fraud detection rate. It was increased from 65.3 percentage exceeding industry average of 86.7 percentage, and the fall positive rate decreased from to 3.7 percentage better than industry average of 18.2 percentage. And the detection time is close to 3.2 seconds, which reduced from 27.4 hours, enabling intervention before funds left in 94.3 percentage of cases. And finally, the annual fraud retention by 19.7 million. That represents a 78.8 percentage improvement with 4 7 4 37 percentage ROA in the first year. Implementation challenges. We have three main, three major areas where we face some of the challenges. The first being data quality issues. Legacy infrastructure comprised 27 distinct systems with 14 different data formats 23 timestamp conventions, and nine incompatible transaction classification schemes. Required 2 83 normalization rules and 1 42 data quality checks to achieve 99.9 percentage data accuracy. And the other one is the latency management. Balancing processing time with detection accuracy. Required sophisticated parallel processing across 1, 2 6 nodes, performing, profiling identified 23 critical bottleneck, reducing evaluation from four. Milliseconds to 1 73 0.4 milliseconds while also maintaining accuracy. Model explainability regulatory requirements across 17 jurisdictions demanded clear justifications for transaction declines. Custom explainability framework generated human readable explanations with 97.3 percentage accuracy in reflecting underlying model logic. This slide highlights some of the key technical innovations. We have adaptive feature engineering. We also have federated learning. And the other one is grab based network analysis. In the adaptive feature engineering we continuously generated new features based on evolving transaction patterns, evaluating. 70,000 4 23 potential indicators monthly. And also identifying 37.2 significant new features automatically integrated into de detection modules. Federated learning, enable collaborative learning across seven participating institutions. Increasing fraud. Example database from two 70 3004 nine to 1.47. While also maintaining strict data privacy through secure multi-party computation. Lastly, we have a graph based network analysis within the key technical innovations. We have marked relationships between four 7.2 million nodes connected by 83.7 million edges identifying 34 previously undetected fraud rings involving one three accounts and 12.3 million in attempted. Fraudulent transactions, Future directions. There are four core areas for the future directions. The first one being behavioral biometrics integration of typing patterns, device handling and interactions patterns to achieve 94.3 percentage accuracy in distribution. Legitimate users from IM posters. Also expected to detect 92.7 percentage of account takeover attempts with just 0.37 false postures. The second core area is voice pattern analysis. Natural language processing to analyze 73 voice characteristics for detecting social engineering attempts in customer service channels with 87.3 percentage identification rate of social engineering attempts. The third one being cross channel correlation. Enhanced detection of fraud patterns spanning multiple channels by analyzing transaction sequences across an average of 3.7 different channels per customer, identifying 42.7 percentage more sophisticated fraud attempt. And lastly, we have quantum resistant crypto implementation of NISG standardized algorithms, including crystals, cyber, and. Lithium to protect against future quantum computing threats to current encryption methods. With this we have come to, and and for the presentation this is what I have for a and h fraud detection in financial services. I would like to thank once again the comp 42 machine learning. Giving me this opportunity to connect with you all and hope to see you and talk you again in the future. Thank you all.