PA Perva. I'm a strategy leader in the finance industry, and today I'll be talking to you about scaling AI and finance. With the focus on ML ops strategies, which are behind the 70.86 billion market transformation, we focus on how production rate ML ops infrastructure is enabling financial institutions to deploy reliable, compliant, and profitable AI and systems in the world. Most regulated industry. Let's quickly talk through the financial revolution and the detailed numbers behind it. 85% of the financial institutions are deploying AI systems today, transforming everything from risk assessment to customer service. 78.6, a number I previously mentioned is the market size. Buy 2030 for this industry. This is the projected value of AI applications in financial services representing massive investment in infrastructure. Thousand plus. That's a big number, but those are the credit variables that are processed in real time by modern financial ML systems for decision making, and then millions of daily transactions happen, and those are analyzed by fraud detection AI with required high accuracy and minimal latency. Latency just meets the time required to do those. As you can see then financial institutions aren't just experimenting with ai. They're deploying at scale and then driving a fundamental transformation of the industry core operations. For today's agenda, we will focus on ML ops of financial industry. There'll be three things that we'll cover. One unique challenges in financial ai, which are regulatory compliance. Model explainability data security and the real time processing requirements that distinguish finance from the other machine learning domains. Second, we'll talk about the production gate machine learning operations architecture, which includes the technical components required for scalable, reliable AI systems in banking, in short and FinTech environments. And then lastly, we talked through some actionable strategies, which are practical frameworks and approaches for transitioning from prototype to production Powerhouse. What are the unique challenges of AI in financial industries? There are four different pieces that we should, we cover one regulatory compliance. All models must satisfy the regulatory requirements, primarily F-C-R-A-E, C-O-A-G-D-P-R, and the basal standards while providing full audit trails. Then there's explain explainability requirements where customers and regulators are entitled to understand the decision factors that go behind the ML models, including like adverse action notices, how this comes into play. Then there's market volatility models, experience rapid trans performance degradation during economic shifts. And lastly, there's adversarial threats where fraud detection systems often face sophisticated attackers actively working to circumvent these detection. So how does financial ML ops differ from the standard practices? There's unique demands of the financial service industry, which requires specialized ML lops capabilities that go far beyond standard ML engineering practices. The standard machine learning operations includes a basic model versioning, has simple AB testing, has general purpose feature stores. Has standard CI CD pipelines, and then basic monitoring dashboards as well as a general on purpose infrastructure. We can be applicable across different industries and places. However, for financial ops, given the challenges we've discussed earlier, we need immutable audit trails with regulatory metadata. We need a champion challenger with segment isolation. We need time series optimized features, stores with point in time correction. We need compliance, integrated deployment workflows. We need segment level performance monitoring with diff detection. We need secure isolated infrastructure with control access patterns. As you can see, financial institutions that attempt to implement AI with it without sub specialized lops infrastructure will face regulatory challenges, model failures, and often security vulnerable vulnerabilities if they're using standard mops. Let's then talk through what is required for machine learning operations architecture for the financial services industry. A robust financial MLS platform integrates specialized components to address the unique challenges of the industry while enabling reliable, scalable AI operations. Each component must be designed with regulatory compliance, security, and audibility as a foundation requirement rather than afterthoughts. What are the components? A big chunk of that is the data foundation. You want your data to have three different kind of qualities. One, it should be a financial feature stop where that, what that means is a specialized four time data. With point in time correctness, which guarantees critical for accurate back testing and regulatory compliance. It needs to support thousands of variables per customer version, feature definitions with lineage, and then batch and online serving capabilities. You also need to make sure that you have a secure data access layer, which means you have role-based controls with fine brain permissions and comprehensive audit logging. This should follow the data minimization principles. It needs to make sure the personal information is handed via a tokenization so no one has access to PII information. You need to have access. You need to ensure there's encryption addressed and in transit. And then lastly, there's a data quality framework. You wanna make sure the data is automatically validated with strict schema enforcement and anomaly direction. You need to have statistical profile monitoring data, drift detection, integrity constraints, validation. The second component of the ML lops is the model development. For that, you need regulatory compliant experimentation, tracking frameworks that capture all model development activities with required regulatory metadata, fair lending analysis, integration, disparate impact assessment model cards with regulatory context. So then as previously mentioned, you need to make sure there's explainability toolkit where you have a pre-approved method for generating customer and regulatory facing information, which relates to adverse action code generation, shap and Lyme integration, as well as counterfactual explanation systems. There are four stages now to the deployment and operations. One is the validation gateway. You ensure this pre-deployment verification, which ensures that the model meets performance, fairness, and compliance standards. Then within that, you have to make sure you have the champion challenge analysis, stress testing across market scenarios, sensitivity analysis on key segments. The second is compliance integrated CI c. You wanna make sure you have the deployment pipelines, have the required approvals, documentation, and the various validation steps. You wanna ensure there is immutable audit, trail of deployments. Role-based deployment controls, automated regulatory documentation. Thirdly, you wanna make sure there segment of their monitoring so you ensure at a segment level there's performance tracking, especially for critical customer segments. They're alert with alerting for regulatory constraint concerns. Within that, you need to have granular performance dashboards, drift detection by segment, and then outlier analysis for high value transactions. Lastly is the automated retraining. You wanna make sure you have a scheduled or trigger based retraining with validation guard rates, so that when there's like a market volatility of our scheduling, like a COVID-19 scenario, you get a trigger. You also wanna make sure there's validation, gated promotions, and then training data set versioning so that after every figure, your training data sets get updated. As you can see each stage, you need to enforce compliance requirements, and this has to be done while enabling operational efficiency. Lastly, for the implementation strategy, you wanna make sure you have a phased approach against. There are four phases. The phase one is the foundation. Where you build core infrastructure, which is focused on data quality, governance, and basic model tracking. Within that, you wanna make sure you have imple, you implement feature stored with regulatory compliance built in, establish experiment tracking with required documentation, create secure development environments with appropriate control in phase two, which is a production pipeline. You will develop automated workflows for work model training, validation, and deployment. Within that, you will have to build a compliance integrated CI ICD pipeline, implement validation gateways with regulatory checks, create model registry with approved workflows, and then coming to the third part, which is operational excellence, where you wanna make sure not only is your model compliant and ready to go, but it is actually contributing to the firm. You will establish comprehensive monitoring and automated retraining, deploy segment aware performance monitoring, implement drift protection and alerting, and then create automated retraining pipelines of validation gates. In the last is where you take a model from not just being good, but to being an industrial leader. You focus on the advanced capabilities, you add sophisticated features for optimization and scaling. Within that, you focus on the imp, you will implement multi-arm banded systems for model selection, deploy shadow mode testing for new models, and then create adaptive monitoring thresholds by segment so that you can constantly evolve and upgrade your model. We've talked through some of this previously, but I wanna highlight what are the common pitfalls and how you can avoid them as you enter the space of ML ops for your financial institution. Don't underestimate the regulatory requirements. Often the problem is that firms discover compliance gaps late in the development, which forces expensive rework. The solution is to engage compliance teams from day one and then build regulatory requirements in technical specifications. Secondly we've talked about this a couple of times now, but there is a concern about inadequate explainability. The problem often arises when the firms are unable to provide required explanations for model decisions in regulatory timeframes. The solution, simple integrate explainability methods during model development and not as an afterthought. So if you start thinking during the model itself, how Q can explain it up to regulators and customers. Thirdly, which is the insufficient segment. Monitoring it basically means that if. Your model's working well overall, but you might be missing degradation in critical customer segments. Despite the overall good performance of the model, the solution is to implement granular monitoring across demographic, behavioral, and product segments. Lastly, it's the poor handling of market volatility. Sometimes models are made for steady state markets, but off, but, and they collapse during economic shifts like a COVID-19 and lead to poor performances. So the solution is to implement stress testing across historical scenarios and ensure that you have responsive retraining triggers. With that, let me. Recall all the pieces that we've talked about and summarize that into the four major takeaways to build production grade financial ML lops. One, compliance is infrastructure. Your compliance underpins everything, so you wanna build regulatory requirements directly into the ML lops components rather than adding them later. Second. Segment level, everything. While your model might look good overall, you wanna make sure you're designing all systems to operate at the segment level for deployment, development, and monitoring, and not as an afterthought. Third, explainability by design. You want to integrate explanation systems from the beginning of the development process so that when your model is complete, you are able to answer any questions related to that. And then lastly, resilient architecture. You wanna build systems that can mean performance through market volatility and data shifts. These are the four underlying pillars of how you would build an ML model, ML lops model. Anything else? Thank you.