Hello everyone and thank you for joining my session. My name is Krishna Uluru, and today I'll be talking about how we built a production ready ML ops pipeline for telecommunications focusing on a neuro network based over IP monitoring system. This work demonstrates how AI driven monitoring can transform carrier grade wise over IP operations. With both high accuracy and automated deployment pipelines. By the end of the session, I hope I will. You will see how integrating ML ops with telecom can make networks more resilient, cost effective, and scalable here did. Here is what I will cover today. First. I will describe the challenges with traditional telecom monitoring. Then I will introduce our ML lops architecture, followed by a closer look at the neural network model we developed. Next, I will walk through the deployment strategy showing how we ensure reliability and scalability in production. Finally, I will share the business results and the roadmap. For ongoing improvements, traditional monitoring has gaps. So let's start with the current state of white monitoring. Traditional tools only provide about 40% visibility into the network, leaving blind spots where critical issues can go Unnoticed. They're also reactive. It typically takes 15 to 30 minutes to detect and respond to issues in telecom. That delay can mean thousands of drop calls and frustrated customers. And finally, the statistical models used for quality predictions only reach around 65% accuracy. This gap is exactly why we need an AI driven approach. Capable of understanding complex patterns and operating in real time. ML Lops Architecture Overview our architecture brings together real time data ingestion, feature engineering, automated training, deployment, and monitoring into a single production ready pipeline. We use Kafka for streaming wise or IP events. Cube flow for training, orchestration, ml, flow for model registry and monitoring through Prometheus and Grafana. The key here is that all of these components are tied together, so models aren't just trained. They are continuously deployed, absorbed, and improved. Let me highlight four essential components. First realtime data ingestion with Kafka processes thousands of events per second at sub 100 milliseconds latency. Second, the automated training pipeline orchestrated by cube flow, which runs distributed GPU based training with hyper parameter training. Third deployment automation. We use blue green deployment models for zero downtime and instant rollback if performance degrades. And fourth comprehensive monitoring where custom metrics and dashboards give visibility into both deployment, both model performance and system health. Together these ensure a carrier grade reproducible and resilient ML lifecycle. Our neural network combines CNN layers to extract features from packet sequences and LSTM layers to recognize temporal patterns across sessions, we added innovations like custom emitting layer for protocol specific features and attention mechanism to focus on anomalies and hierarchical feature extractor across timescales. The result, 88% prediction accuracy with a 23 point movement over traditional statistical methods. That's a major leap in how accurately we can predict and preempt call quality issues. The model is only as good as its features, so we built a fully automated feature pipeline and the edge high throughput collectors capture sip, RTP, and QS metrics. These are processed in real time with Apache Beam handling over 10,000 events per second while extracting 45 plus features. We use recursive feature elimination for automatic feature selection and feast as a feature store to ensure consistency across training and inference. All of this is logged and monitored to maintain data quality and lineage. Training is orchestrated through cube flow pipelines. We use TensorFlow data valuation for schema checks and drift detection. Then run distributed GPU training at scale. Every experiment, parameters, metrics, artifacts, is tracked in ML flow, so results are reproducible. Models are versioned and registered with clear approval workflows so teams can roll forward or backward confidently in production system maintains 96% anomaly detection accuracy while handling thousands of concurrent requests with 99.99% available. We achieve this with containerized models. Intensive flow serving orchestrated by Kubernetes with horizontal auto-scaling. Bluegreen deployments provide zero downtime and edge inference Nodes across regions deliver responses in under 50 milliseconds. This is a true carrier grade ML deployment. Comprehensive monitoring system has been de developed. Monitoring goes beyond just accuracy. We track model performance metrics like precision and recall data quality metrics like feature rift and missing values, system health such as latency and throughput. When drift is detected, retraining is automatically triggered. We also run. AB tests in production and use self-healing mechanisms for infrastructure issues. This makes the system adaptive, not just reactive. Trust is critical in telecom operations, so we have built explainability into the pipeline. We use shop values for feature importance. Counterfactual explanations to show how outcomes could change and dashboards that track feature attribution over time. Automated model cards and audit logs document every version and decision. This means engineers can always trace a prediction back to its root cause, essential for both regulatory compliance and operator trust. What does all this mean in practice with this system fault detection time has dropped from 30 to 15 to 30 minutes, down to two to five minutes. Network visibility is greatly improved. Licensing costs are reduced through open source tooling and uptime has reached carrier grade levels. Most importantly, customer satisfaction improved because quality issues are caught before they affect end users. We rolled this out in phases. Phase one, the foundation, the data collection, feature engineering, and an initial model, phase two. Is automation CICD, pipelines, testing registries and dashboards. Phase three is scale distributed training and geographic deployments. Phase four, ongoing optimization, ensembles tuning and knowledge transfer. This phased approach allowed us to deliver incremental value while building towards the full vision. To wrap up three points. First, production ready ml requires end-to-end thinking from data to monitoring. Second, automation drives reliability, especially at telecom scale. And third, explainability builds trust, which is essential for adoption in mission critical systems. These principles. Guided us in creating a system that's not just research project, but a reliable production grade ML ops system. Thank you for your time. I hope this session gave you insight into how ML Ops can revolutionize telecom monitoring. Thank you.