Hi everyone. Thank you for joining. My name is Bin Aja and I'm a DevOps lead within IBM. Today. Today we are going to talk about ethical AI deployment at scale. And like we are going to discuss about the platform engineering approaches. So today we are going to dive into one of the most critical challenges in modern tech which is ethical AI deployment at scale. This is not the theoretical discussion. It is like a hard engineering problem, which we are facing every day. And the key to solving this problem is to, it lies with us like platform engineers. So our, over the few minutes we will move from understanding the core dilemma we are facing to the tangible ROI of solving it. And finally we'll be talking about like a practical blueprint for building ethical safeguards, which is directly into the fabric of our AI platforms. So let's get started. So today we are seeing is 83% of enterprise accelerate AI integration and platform team are facing more lot of pressure to deploy AI responsibly and maintaining. The velocity as well. So this session, how we are embedding like ethical AI safeguards into the deployment process and so that we can create the pipeline faster and most lively. And so that AI systems so that we can the stakeholder trust, right? So now we have a platform engineer dilemma, right? The dilemma we all recognize, like we are caught between a two. Powerful challenges which seems to be like opposing forces, right? On the one hand we are facing like the velocity pressure and so that we need to ship the product faster and support more models. And on the other hand the other hand, we have a non-negotiable mandate for responsible ai. So we must embed like a fairness. Avoid biases and protect the privacy as well. And also and also ensuring the compliance, right? So the consequences of getting things wrong are severe and it, and like a reputational damage is there. And there are regulatory fines and also a total loss of user trust if things go wrong. So the old mantra used to be like. We can go fast and break things, but in today's world, breaking things, breaking people. So the challenge here is like, how do we maintain a balance between a breakneck speed without breaking our commitments? So now we are going to talk about the ROI of ethical AI platforms. The good news is that it is not just a cost center. Ethical AI platforms building these platforms deliver like a massive and measurable return on investments. So think about it like what happens when you automate ethical checks. First you accelerate the velocity. That is the benefit we get. And like by catching issues early in the pipeline, we avoid the mass massive delays. And also we avoid the rework. Required when problems are found late in the production. So it leads to it leads to the improvement in the deployment efficiency. So we see two 86% improvement in deployment efficiency, right? And the second one is like we are de-risking the deployment. So automated testing is leading to 89% detection rate for bias issues pre-production and 84% enhancement and compliance. So that is a big achievement. This is not just about avoiding fines. It is more about getting more user trust and this trust translate into adoption, right? So as we can see like organizations are like there, there's a 92% increase in sustainable AI adoption these days. And because developers are trusting the platform and end user is trusting the outward. So this is how ethics becomes, right? A competitive advantage, not a bottleneck. So that's why it is important that we integrate ethics. Into the picture. So now we will talk about automated ethical testing in the next slide in the pipelines. So how do we achieve it? It starts by shifting left. Ethical testing must be automated and integrated into the pipelines. This means treating and ethical failures like a build failure. Or a unit test failure. So as soon as a model is committed, the pipeline should be automatically run it against the battery of test. And we should check for a statistical parity differences and equal opportunity. And and also using using tools like AI Fairness 360 or lan, we can achieve that. And the key here is the process should be automate and mandatory. It is like a quality gate. If the model does not pass this gate and the bias check the pill should fail automatically. And this moves ethical validation from manual to automated validation and so that we can automate this week long processes and we can and we can also reduce validation cycle by 55%. Oh, like close to 56% and 91% rapid issues. And it all mitigation effectiveness also, and this is this is the foundation here and making like ethics is like a non-negotiable part of the definition of done. So now we will talk about talking about the platform native vehicle frameworks. If you talk about the tools, so tools are not just enough, right? We need to build a, like a platform framework which means like a big banking ethics into the infrastructure we provide. So we need to bake ethical frameworks into it. So instead of asking every development team to figure out on their own, if we can, if the platform team provides like a curated, pre-approved ethical, tooling as a service we offer. So what we can do is we can standardize model cards, which will auto-populate the fairness metrics, right? Second one we can do is we can pre-build compliance data pipelines with the anonymization building, right? Third one is we can provide the approved libraries for bias mitigation, right? So provide, by providing these, services like this as like a managed services, we ensure like a consistency into the product and we reduce the cognitive load for product teams. And also we guarantee every model deployed into our platform meets like baseline ethical standard by default. So we make the right way and we make the things right way and in the easy way. Now we are going to talk about the case study for a healthcare AI platform. So let's make like a concrete healthcare study discussion. So the stake here could be could not be higher. We are dealing with a lot of patient data and health data, so the platform team implemented these one of the platform team implemented these exact principles, which we just talked about, and they embedded like a ethical testing for bias in diagnostic models. And built the privacy preserving data workflow directly into the platform. And and it gives 67% reduction in the privacy incidents, and it gives like a 79% increase in clinical trust. And 92% sustainable a adoption and 74% fewer bias concern, which is a big achievement. Now so the doctor could focus, on a patient care because they trusted the platform more and at time ensured that the model was fair and data was safe, right? So now we are going to talk about the cloud scale ethical monitoring. By implementing monitoring in place, it helps a lot. So how, so deployment is not just about finishing the line, right? It is more about how we can, how model can. Like model can degrade in the real world through model drift and data drift, right? So what was fair yesterday might not be the, might not be fair today. This is where cloud scale ethical monitoring helps. So we need to monitor pipelines. So we need to monitor productions in the real time and like by. Like by monitoring like real time events which has probably been processed like thousands of events per second, right? And checking for drift and and violations against our ethical baselines. So we should be monitoring all of these. And for that we need a dedicated monitoring stack. So perhaps using tools like Fiddler. Or s core which like the, we can grade, we can achieve the great outcomes. Like the outcome is to detect 91% violations we can detect and it'll improve the response time. Like we can increase, like you can get a 78% faster incident response time with the help of this. So like automated alerting and remediation workflows, engage the right teams immediately whenever the potential issues arise. So now a critical question is like, who decide? What is fair, right? So the, it cannot be like an emerging decision made in a vacuum. Like what, who decide what is fair? This is where the cross-functional governance is very much essential. We need to build a, like a lightweight council with legal compliance, ethics and business, right? With the business representatives. The job here is to define what our. What are like, what are the fairness thresholds? What is our definition of bias? So our job as a platform engineer is to encode that. Into the policy as a code and so that we can turn their human decisions into automated gates in our pipeline. Like that's why we are saying like a policy as a code and the CO and this kind of collaboration which will make the system more scalable and estimate, and it is why organization using this. Model report I'll get 75% better project success rate and 91% effectiveness in the mitigating issues. So now it is very important, like we build our ethical a platform. So how do we do that, right? So for that we need to define the platform requirements. We need to integrate ethics into the CICD pipelines. And then we have to develop operation monitoring, right? So first we need to, pick one high visibility model and audit it and create audit it for a single metric like gender bias. And then tool up, like experiment with the. Then we do experiment with one open source framework and run it manually or model and see what it finds. Then we can third, integrate it like and add like a single ethical test for a non-blocking check into our pipeline and let developers see the report. And then fourth, like we can mandate like we can graduate the test like a blocking it for our critical applications, right? And then we can add more tasks into it. And then we can develop operational monitoring into it and formalize our governance council which is not a journey. It is like a journey, not a one stop. Like another flip you with the project, right? So it is, it takes time to develop the monitoring in place. So the key takeaway here is indeed. Ethical AI into it. Don't append. So ethical AI s should be more infrastructure components or not. It should not be an afterthought or add on to an existing platform. And we should automate the validation building automated ethical AI testing, like ethical testing into CICP pipelines to catch issues early without slowing deployment. And that is like govern through platform. So implement like a platform native governance, which is scales with our AI initiatives. And it was through like with the regulatory requirements. So when it is done right, the ethical ai framework don't slow the deployment. It accelerate the sustainable AI option by building confidence into the system and increasing the user trust. So thank you for attending and please let me know if you have any questions.