Rohit Kandu developed an On-Demand Profiling feature for the NVIDIA/Megatron-LM repository, enabling dynamic inspection of training workloads through a new command-line interface flag. He integrated the profiling server’s startup logic directly into the training script, allowing users to activate workload inspection without modifying code. This approach improved observability and accelerated performance debugging for large-scale distributed systems. Working primarily in Python, Rohit applied his expertise in system configuration and command-line interfaces to streamline the profiling process. The feature laid the foundation for future profiling-driven optimizations, demonstrating thoughtful engineering depth in addressing the challenges of monitoring and tuning complex training workflows.