Hi everyone. My name is Bar. I'm currently working as senior Salesforce consultant with over nine years of experience in Salesforce and cloud technologies. Over these many years, I have delivered Salesforce implementations, cloud integrations, and AI driven solutions across industries such as finance, healthcare, and registry. In today's session, I will walk you through the topic, architecting AI native platforms. This talk will focus on how infrastructure, workloads, and governance come together to make AI work at scale. We'll also look at real world implementation patterns and future threats. By the end of this session, you will have a clear understanding of the challenges, best practices, and takeaways. For building scalable AI native platforms, the infrastructure imperative, the starting point for AI initiative is infrastructure. Without the right infrastructure, most AI projects remain in experimental phases and fail to reach production. AI requires compute intensive gps. High performance networking and scalable storage. There are two risks to highlight here. First, organization that embrace infrastructure can achieve a competitive advantage, deploying AI models faster and more reliably. Second, those that focus on experiments without production. Ready infrastructure fall into what I call the experimental cloud. This means AI never skills beyond pilots. So the message is clear. Create infrastructure as the foundation of your ai. Understanding AI workload characteristics. A workloads are very different from traditional applications. Training workloads require massive parallel processing where GPU Shine inference workloads, on the other hand, focus on speed and scalability. Think of recommendation engines, chart bots, fraud detection models. They need millisecond responses. This means infrastructure must. Consider both sides. Training environments that are compute heavy and inference environment that are lightweight, but scalable. If we don't understand workload characteristics, we risk over provisioning, under utilization and spiraling costs. Data pipeline, I. Data is the fuel of ai without well-structured and governed data pipelines, even the most sophisticated models will fit a robust data pipeline. Ensures data is collected, lean, transformed, and served consistent. Future stores are becoming a critical component. They allow teams to reuse features across models. Ensure data consists stream processing. Add another dimension, real time insights. For example, in fraud detection, if data pipelines lack even by a few cycles, the opportunity to stop fraud is lost. Infrastructure patterns for scalable model trade. Scaling model training is not driver. We face challenges in resource management, container isolation, and network topology. Kubernetes and container orchestration solve many of these challenges, but GPU scheduling, network optimization and distributed training strategies are equally important. For example, let's say distributed data parallelism allow us to train large models across the multiple GPUs while tolerance mechanism ensure that even if one not fail, training continues the seamless. Resource management challenges. Managing resources is a balancing act. The key challenges include allocating GPU and CPU resources effectively. Isolating workloads through containerization. Designing network topologies that support high data building system T to hardware failure. If resource management is not handled properly, fast train rocket and innovation slows. That's why AI n native two platform. Must embed resource governance from day GPU and CPU optimization. AI workloads often involve a mix of CPUs and GPUs. CPUs are great for general purposes tasks while GPUs accelerate metrics, heavy competitions in model training and inference. The challenge is in optimizing utilization of both eternal genius resources. Scheduling ensures that workload are directed to the right compute layer. This improves efficiency and lowers cost. For instance, lightweight reprocessing me run on CPUs while the heavy model training runs on gpu. Data pipeline architectures, modern AI platforms rely on advanced data pipelines, a future store, ensure that right data is consistently available for training and inference frame processing, handles real time data, allowing systems to react immediately. The architecture must balance batch and real-time processing. Batch pipelines are great for large historical data sets while stream pipelines ensure responsiveness. Together they provide a comprehensive data bone backbone for ai YouTube platforms. Observability and monitoring for ai. Unlike traditional systems, AI platform requires observability and multiple levels, infrastructure, data, and models. We need to monitor GPU, utilization, pipeline Health, and most importantly, model performance. Metrics such as accuracy, precision, recall, fairness indicator must be continuously tried. Observability help us detect entire drift bias and performance degradation early, preventing the costly failures, real world implementation patterns. AI native platform. Looks very different across industries. Some examples, high frequency trading. It requires ultra load latency, infrastructure, and real time pipelines. Content recommendation, it needs scalable system that personalized the individual level. Healthcare here. It particularly focuses on compliance, accuracy, and ethical safeguards. These industry specific examples demand us that there is no one size fits at the level of AI platform Architecture must align with business goals and complex needs. Performance optimization. Performance optimization is about making sure resources are used efficient. These involves optimizing GBU memory usage, improving storage performance with MVM and tuning models, serving layers to reduce latency. For example, batching interface request can reduce overhead while catching the frequently accessed futures improve. Small optimizations at scale create massive performance scales. Security and governance in AI platform in AI platforms cannot succeed without robust security and governance. These include. Protecting models from ARISAL attacks, ensuring data privacy and implementing role-based access. Governance also means ethical oversight, making sure AI decisions are transparent, explainable and T with regulations such as GDPR and hiphop. Future trends in AI infrastructure. Looking here, several trends will shape AI native platforms, each AI bringing intelligence closer to the source of that quantum computing, unlocking optimization problems. Traditional hardware struggles with Pneumo, morphic computing. Mimicking the human brain for efficiency. Automated EML and explainable AI democratizing ai while keeping it transparent. These trends will transform the way we think about scale and adoption. Building the roadmap for the foundation of AI driven innovation. Implement AI native platform successfully. Organizations particularly need a roadmap. These include short term wins, such as building feature tools, medium term goals, like integrating observability and long-term strategies like adopting hr, quantum compute. The roadmap should balance innovation with governance, ensuring sustainable AI adoption. Key takeaways and the conclusion. Coning AI native platform is able uniting infrastructure, workloads, and governance. Here are the key takeaways. Infrastructure is the foundation of scalable. Understanding workload characteristics for effective design data pipelines are the backbone of reliable ai. Observability ensures continuous implement governance, build trust and compliance future trends, demand ongoing division. Thank you for joining this session. I hope this helps you. Strategically about building a to platform. Thank you.