Good afternoon everyone. My name is She Gupta, and I'm thrilled to be here today to talk about a truly transformative trend in the world of artificial intelligence. Serverless AI rise revolutionizing ML deployment with scalability and cost efficiency. Over the next 25 to 30 minutes, we'll explore how serverless computing is not just an incremental improvement. But a paradigm shift in how we manage and deploy AI solutions. We'll see how it is enabling organizations to build and scale AI's application faster, more efficiently, and often more cost efficiently than ever before. As you can see, serverless computing has already made a significant impact offering automatic scaling. And paper use models without the headache of managing underlying infrastructure. The market itself valued at USD 17.2 billion this year is projected to grow significantly hitting A-C-H-E-R of 14.1% through 2030. Okay. This presentation will delve into how serverless AI is streamlining deployment, boosting resource utilization, and ultimately accelerating innovation. Let's begin. So what exactly is serverless AI architecture at its score? It operates on an event driven model. This means computing resources are dynamically allocated only when needed in response to specific triggers. Think of it like a light switch. The power is only on when you flip the switch. The key components as illustrated here include even sources, function, containers, supporting services, and finally, resource management for even sources. These are the triggers I mentioned, things like API request. Changes in data or schedule task, they initiate the process function. Containers. These are stateless environments where your AI model code actually executes. Stateless is crucial here. Each function in vocation is independent, which is what allows for massive parallelization and seamless automatic scaling. Third, supporting services. These are essential backend components. Authentication data, storage monitoring, all managed for you. And finally, resource management. The platform handles the dynamic allocation of resources based on the workload demands. This architecture significantly reduces the burden of infrastructure management. In fact, organizations leveraging this approach have reported up to a 68% in the reduction in infrastructure management overhead. That's a huge win. Freeing up teams to focus on building great ai, not managing servers. Now we understand the architecture. Let's look at how serverless AI solutions are typically implemented. It generally follows a cyclical pattern as you can see in the diagram model. Encapsulation first developer package, their machine learning models within their stateless functions. These functions are often triggered via STTP request for inference points. Or other events. The beauty here is the minimal IV infrastructure code, second deployment. Next, this packaged function, which was created in model encapsulation, is pushed to a serverless platform along with its trigger configuration. Third execution. When an event occurs, say an API call with new data, the function runs performing the AI task, like making a prediction. Scaling and critically, the platform automatically allocates resource based on the demand. If you see a sudden spike in request, it scales up. When demands drop, it scales down. You only pay for what you use. Modern serverless platforms are quite versatile, supporting diverse simul frameworks and model architectures. We are seeing successful deployments of complex deep learning models ranging from 75 MB up to 1.8 gb, all while maintaining impressive performance, like an average inference time of just 1 1 2 milliseconds. Can you believe that? Okay, so of course not all serverless platforms are created equal. Choosing the right one is critical for AI deployments, especially when considering performance, cost, and specific functionalities like GPU support. This slide presents a snapshot from standardized benchmarking studies that used identical deep learning models. RESNET 50 birth base, and YOLO V four across major cloud providers. Let's talk about GCP Google Cloud Functions, which is a product of GCP standouts for an image classification workloads demonstrating the lowest average inference latency at 1 1 2 milliseconds, and also the fastest cold start time. Their GPU support is also quite comprehensive for Amazon. AWS Lambda follows with a strong 1 36 millisecond inference latency and their Snap Start feature. Significantly improved Cold start time for Java functions with good but limited GPU support. Currently it's on T four. Azure functions for Microsoft offers the most consistent cloud Cold Start performance and container based GPU supports. Last but not the least, A BM Cloud functions while cap. A capable platform currently shows slower inference and cold start times in these benchmarks and lacking direct GP support for function. Let's see, what's the key takeaway here? It is the cold start, the initial delay. When a function is invoked after a period of inactivity, it is called a cold start. This could be a significant factor, especially for A GPU accelerated functions. Google Cloud functions, for instance, showed 58.7% faster initialization time for these. So depending upon your specific AI workloads, sensitivity to latency. And the choice of platform matters greatly. Okay, so now let's talk about the key advantages of serverless ai. The benefit of adopting serverless AI are compelling and transformative. Let's look at some numbers here. It has 64.8% deployment time production. Imagine cutting your deployment model cycles from an 85 hertz down to approximately 29 hours. This is a reality for many teams adopting serverless, a massive improvement compared to a traditional infrastructure. 79.3% resource utilization improvement. Serverless ensures you are using computing resources much more efficiently. No more over provisioning services servers that said ile, most of the time, 38.4% cost reduction. This directly translates to average savings compared to the traditional deployments. As you are only paying for the compute time, you actually consume. 99.95% deployment availability. This high reliability is inherent in the managed nature of the serverless platform. So to conclude, as mentioned earlier, automatic scaling capabilities are a game changer for various AI workloads. Platform can scale from handling 40 requests per second to over 4,200 requests per second in just over three seconds, all while maintaining the response time of under 23, 2 35 milliseconds. This agility is incredibly valuable. Okay, so let's make this entire numbers concrete with a real world example from the financial services industry. A leading global investment banking was struggling with the scalability for their traditional risk analysis infrastructure. They had 38 dedicated high performance servers, but they were operating at an average utilization of only 31.7% A. So it's a classic case of over provisioning and inefficiency. They decided to migrate to a serverless AI solution. The migration took about 4.5 months. And it wasn't without challenges adapting to complex workloads and ensuring regulatory compliance were key hurdles, but the post implementation results were substantial. 68.2% reduction in infrastructure cost a direct result of optimized resource utilization. Three one 5% increase in the peak processing capacity they could now handle over. 12,000 concurrent model execution. During the peak reporting periods, something unimaginable before 62.2% improvement in the processing speed. Average latency for a risk analysis dropped from 8.2 seconds to 3.1 seconds, and crucially for finance enhanced regulatory compliance through comprehensive audit trails and improved security protocols inherent in the new system. This case study clearly demonstrates the tangible business benefits of serverless AI in a demanding regulated environment. Now, let's take a real life case study for another critical center, which is healthcare. A healthcare technology provider specializing in medical diagnostic tool faced similar challenge their traditional server based architecture for a medical image processing required. Significant upfront capital investment, specialized personal, and suffered from a long infrastructure provisioning timelines, averaging 86 days to expand service. They adopted a serverless AI solution using container I achieve learning models. The migration spanned over seven months, over three phases with a strong focus on security, strict access controls, and comprehensive encryption. The outcomes were impressive. First operational improvements, a 73.5% reduction in the operational overhead. Their auto-scaling capabilities now support variable workloads from 50 to 5,000 images and R without any manual intervention deployment efficiency. Their deployment cycle for a new model updates plummeted from 36 days to just 4.8 days. This allowed them to rapidly interior research advances into their production system. Financial impact, they achieved a 41.3% reduction in total expenses despite a two one 2% increase in the processing volume. Imagine that. Number furthermore, the granular price pricing models of serverless server allowed them to better serve smaller healthcare providers. Again, we are seeing serverless AI delivering significant improvement in efficiency speed. Cost even while handling increased demands. Okay, so now let's talk about innovation and rapid development and how it helps there. Beyond operational efficiencies, a serverless paradigm fundamentally changes how organizations approach innovation and rapid prototyping in AI development organizations leveraging serverless architecture have on an average, reduced their development cycle time by 57.8%. High performance teams are even seeing up to a 73.5 reduction in time to market for the new AI features. That's incredibly fast. Rapid prototyping in particular, sees and impressive games development teams can now deploy and test a new AI model variance in just 3.4 hours compared to nearly 23 hours in the traditional environment. This agility, flu fuels increased experimentation, reduced complexity, improved collaboration, and accelerated time to market. You can see the numbers here for increased experimentation. Hydration speed can be increased by 2.9 x, allowing for much faster validation of AI models and hypothesis teams have reported a 68.4% reduction in the implementation complexity, which also leads to decrease in your bug density by 52.3% and few post-deployment issues improved. Collaboration code reusability improves significantly with teams reporting a 53.2% improvement. Average feature deployment time dropped from 15 days to just under five days with a high first deployment success rate of nearly 79%. So it is safe to say that serverless empowers AI teams to innovate faster and more freely. Okay, so now let's talk about technical challenges and limitations. It's important to be balanced despite all the advanced advantages which we have discussed so far. Serverless AI isn't what isn't without its technical challenges and limitations. Cold start latency. This remains one of the significant issues. Complex deep learning models can experience significant delays up to 6.2 seconds during initial initiation. While platforms are improving, it is a factor to consider. Memory limitation. There are constraints on the model size and the complexity that can be loaded into a function's memory. Execution time limits functions typically have a maximum execution duration limits ranging from two 50 to eight 50 seconds. This affects nearly 39% of the complex AI processing task that might be longer running. Vendor lockin dependency on provider specific features and services can make it harder to migrate between the cloud members. Imagine you started with GCP and then you wanted to shift to Microsoft in the between determine. Imagine the number of times the number of ads you have to spend a resource constraint while scaling is automatic. The available compute options, your CPU types, memory configurations for individual functions might be more limited compared to dedicated virtual machines. Especially for a very large or specialized model. In fact, over 82% of the enterprise AI deployments encounter at least one resource related constraint. These are real considerations that organization needs to plan for when adopting serverless ai. Okay, now let's talk about security. Security is of course, paramount in any AI implementation. So first less introduces its own unique considerations compared for to the traditional architectures isolation mechanism. Functional isolation is a primary defense against the cross tenant vulnerabilities. Container-based isolations used by many platforms provides effective security boundaries. For a vast majority, nearly 95% of the common attack vendors. Access control, effective identity and access management is fundamental. Around 69% of organizations are implementing fine-grain access control at the functional level, ensuring functions only have permissions they absolutely need. The use of short-lived credentials has been particularly effective, reducing credential misuse incidents by over 72% data protection, end-to-end encryption strategy for data addressed in transit. And during processing are crucial and well supported by serverless platforms compliance. Comprehensive logging and monitoring capabilities are available, which are essential for meeting regulatory requirements and for auditability. While the attack surface ships with serverless robust security practices and platform features can effectively mitigate risks. Okay, so now let's talk about the future outlook. Looking ahead, the serverless AI landscape is set to evolve rapidly. We are expecting significant transformations in the serverless platform with increased specialization and optimization, especially for AI workloads. One exciting concept is nano functions. This represents a ship towards even smaller execution units. Capable of executing specific components of a computational graph rather than the entire model. This could lead to more precision, precise resource allocation, improved parallelization and early results already showing a potential 37.8% reduction in the overall execution time. Here's a potential timeline of emerging trends. Let's quickly go over 25 to 26. We likely see standardized function interfaces and deployment specification emerge, which will help reduce vendor lockin service. Lockin concerns 26 to 27 expect an ultra light weight container format optimized for AI workloads 27 to 28. Predictive scaling algorithms could achieve a very high frequency in workload forecasting. 28 to 30 platforms will very likely incorporate more inbuilt capabilities for responsible ai, such as bias detection, fairness evaluation, and explainability tools. Therefore, the future of serverless AI is bright and full of innovation. Okay, so to recap, serverless AI offers a powerful new way to deploy and scale machine learning models. Bringing significant benefits in terms of cost, efficiency, speed, scalability, and fostering innovations. While challenges exist, the ongoing advancements and a clear value proposition make it an interesting, compelling choice for organizations across industries. Thank you very much for your time and attention. I'd be happy to answer any questions you may have. Thank you.