Worked on enhancing observability for the NVIDIA/TensorRT-LLM repository by implementing Prometheus-based telemetry for inference operations. Developed and integrated new metrics to monitor prompt cache usage, speculative decoding, perplexity calculations, and batch occupancy, all within the Python backend. This instrumentation enabled comprehensive end-to-end monitoring, supporting data-driven performance tuning and more effective capacity planning. The approach focused on seamless integration with existing backend development workflows, leveraging Prometheus for robust metrics collection and performance monitoring. By providing detailed insights into inference behavior, the work facilitated faster incident response and improved reliability, laying the groundwork for ongoing optimization of the TensorRT-LLM framework.