Leading-edge expertise in LLMs, including transformer architecture, attention, tokenization and positional encodings, and how models are trained and served, with advanced PyTorch skills spanning custom model code, mixed precision and distributed training (e.g., FSDP, DeepSpeed). Deep knowledge of inference optimization, including prefill and decode, KV-cache management, continuous batching, speculative decoding and quantization, and serving at scale with vLLM, SGLang or TensorRT-LLM, ideally on AWS and including GPU kernel work (CUDA or Triton).