LLMs and deep learning models actually run on hardware, training versus inference, matrix multiplication, KV-caching Real experience building performance models or forecasting tools, Python or spreadsheet-based, from research, a thesis, a placement, or serious personal projects Hands-on work with GPU or accelerator code, CUDA or similar Familiarity … with profiling tools (Nsight, PyTorch Profiler) and ideally some exposure to monitoring stacks (Prometheus, Grafana) Strong Python for data work, Pandas and NumPy, genuine scripting ability Nice to have: exposure to inference serving frameworks like vLLM, published research, or open source contributions in this space. Why look twice at this ...