Hello everyone. Good morning or good evening, depending on your location. I am TU and I have 23 years of experience in data engineering, AI and a ML. Cloud Technologies, enterprise bi, and in manufacturing industries. It is a pleasure to be here at Platform Engineering Conference. Today I'm going to take you through the journey of building scalable AI ML platform for industrial iot, specifically through a cloud native approach for predictive maintenance infrastructure. This session combines of my 23 years of experience in enterprise school data platform with real world lessons from industrial transformation projects. Let's get started. Here is the agenda for today's conference. The industrial iot landscape. Challenges and opportunities evolving from reactive to predictive maintenance. The business use case for ai, ML powered maintenance core architectural patterns for iot platform success. Cloud native platform components for industrial iot edge computing integrations. ML Pipeline orchestration for predictive maintenance developer experience, enabling data science team case studies for multi-site manufacturing implementation roadmap, and key takeaway for platform engineers. The industrial IOT landscape, challenges and opportunities industrial iot has transformed how manufacturing environment operates. In industrial iot, we see 1.52, 2.3 data bytes of daily sensor data is flowing from distributed sites, each with mix of device. Protocol and requirements. The opportunity here is huge, but the challenges are from unifying heterogeneous data stream to meeting strict uptime requirement. Our mission as a platform engineer is to turn this raw data into actionable intelligence while keeping the enterprise system reliability intact. Evolving from reactive to predictive maintenance. Here we have four types of maintenance, reactive scheduled maintenance, conditional based maintenance. Finally, AI driven predictive maintenance. Let's start with reactive maintenance. Traditionally, reactive maintenance has been fixing the things after they file, which needs more downtime and unplanned production disruptions will cost you more than it came Schedule maintenance. Which is often meant fixing too early or too late, which is also not best practice. Next is conditional base maintenance. Maintenance, which improve the things, but still lack of foresight and some limitations are there in this method as well. The final only is AI driven. Predictive maintenance is the game changer. It's giving you eight to 10. 12 days of advanced warning with almost 90% prediction accuracy. But to get there, we need infrastructures that supports this evolution and helps organizational changes both technologically as well as culturally, the business use case for AI powered predictive maintenance. Here are the game changing results. Downtime reduction. Ines implementing AI driven maintenance have reported 50% reduction in downtime, which is half the machine offline time compared to before return on investment. Perhaps most impressive is 385% return on investment. These projects don't just play themselves, they multiply the value delivered. Next is. Predict prediction accuracy. Model M and models can predict equipment failure with 89%, 89.7% accuracy. This means engineers and operators can get clear, timely, alert. The last is early. Warning alert, these models can provide eight to 12 days of advanced warning, which is critical for planning downtime and avoiding production disruptions. This aren't just nice to have a metrics. This trans, they translate directly into operational and financial win. That is why 60% of industrial manufacturing or investing in predictive maintenance platform. Here. The question is, how do we build them at scale core architectural pattern for iot platform success? It start with architectural choice, multi-tier processing. Use edge device for quick real-time decisions, and then use cloud for deeper and more complex analysis. Sometimes mix of both. Hybrid is the best to balance speed and heavy workloads. Second is even driven architectures instead of waiting in a line, data is handled as even happens. This lets systems smoothly manage challenges, data volumes, and unexpected, sudden surge of many sensors, auto-scaling infrastructures. Kubernetes driven elastic city for peak loads, which means system dynamically can scale up or scale down to match the demand and workload. The last one is distributed storage strategies. This keeps, data is different layer. Hot storage for faster access to real-time A and cold storage for keeping older historical data used in a training and long-term analysis. These patents provide the backbone of the scalable resilience platform, cloud native platform component for industrial iot. From ingestion to association, every layers matters. Let's start with data addition layer. It connects machines and sensor using standard protocol like O-P-C-U-A-M-Q-T, and Modbus and stream streams data with tools like Kafka and puls. Next is storage and processing layer. Source sensor data in time series database and large scale data lakes like Snowflake and Databricks ML orchestration layer. It automates model training with tools like Cube flow or ML flow, and manages reusable features in features stores. The last is application layer. It provide graphical QL, API and rest a PA for data access and powered real time dashboard with alerts. The key is loose coupling so that each layer can evolve independently, each computing integrations. Bringing AI to the source, one of the most powerful enable. Enabler here is edge computing. Here are the some key benefits of edge computing. Critical alerts are handled up to 75 to 80% quicker only important data is sent for reducing network. Network load and bandwidth systems keeps working. Even if the internet connectivity drops, press pre-processing edge. We'll cut down the expense of cloud usage. Next is platform engineer considerations. Deploy ML model efficiently on the device with limited memory and processing power. Update edge device remotely without manual inter intervention. Ensure data is exchanged safely using mutual authentication and secure communication. Keep the data consistent between edge device and the cloud. Even unreliable connections. Even with unreliable connections. The ultimate goal is here, seamless data flow between Edge and the cloud. Cloud platform. ML Pipeline orchestration for predictive maintenance. Predictive maintenance is not just about the models, it is about the full lifecycle development. Start with data collections, gather the real-time mission signals such as vibrations, temperatures, pressure, et cetera, from iot sensor. Second is processing clean, normalize and extract meaningful patterns from raw sensor data. Third is train the ML model. Use historical data to build algorithm that predict occupant failures for this validation against real world failure. Test the model predictions against actual breakdown events to ensure accuracy. Fifth deployment into production. Integrate the model into live system for real time failure. Productions prediction. Last one is monitoring continuous, continuously track the accuracy, data drift and retain when the performance declines. Automation, observability and governance needs to be built. Every stage. Developer experience enabling data science team. Data scientists should focus on the model, not on the plumbing. That means self-service infrastructure give the data scientist on demand access to scalable compute, storage and tools without waiting on operation teams. Next is automated CICD for ml. Streamline the model development with pipeline, that auto test version control and auto deploy ML code. Last one is manage the feature stores. It provides centralized, reusable system for consistent feature engineering across training and production systems. Yes, strong developer experience accelerate adoption and productivity while maintaining the governance. Case studies, multi-site, multi-site manufacturing implementations. Let me bring this to life with some real time use case studies, a global manufacturer with 37 facilities and 12 different equipment vendor, and a major downtime challenges. Our platform solution is to use the hybrid cloud. Which is AWS and Azure architecture, snowflake analytics, cube flow pipeline, and the custom adapters Here are the end results. 43, 42 times 40, 42%. Downtime reduction, 4.3 million. Annual saving, 93% model deployment success rate, and a 2.5 times productivity boost for data science team. Implementation roadmap for platform evolutionary strategy. We approach this in different phases. First, to start with foundations, it takes three to four months of connectivity, building basic pipeline, and data governance. Second ML deployment, it takes two to three months for feature engineering training and initial deployment scale and optimize. This process takes four to six months of edge integration, advanced automation, enterprise integrations. Last one is enterprise. Expansion. This is ongoing process platform as a service offering multi-site ana analytics and continuous evaluations. This phased strategy delivers value early while building towards full maturity. Here are some key for platform engineers, architecture matters, design end-to-end system that simultaneously handle data from sensor to insight, ensuring scalability, resilience, and edge and cloud integrations. Next is developer experience drives adoption, provide a simple and secure abstractions. So that data science scientists can focus on the models without wrestling with infrastructure team. Last one is incremental wins, which means start small scale. Small meaning begin with small, which has a high return on investment use case, and then scale globally using model modular platform that. Adopt to new needs and technologies. The most successful platform aren't just a technical achievement. They are enabler of organizational transformation from reactive to predictive operations. Thank you for time today. I look forward for your questions and discussions about how we can shape the future for predictive maintenance together. See you later. Bye.