Arpit Singh Gautam

Python, CUDA/PTX, Triton, C++, SQL. PyTorch, vLLM, TensorRT-LLM, DeepSpeed, TRL. Quantization (GPTQ, AWQ, SmoothQuant, NF4/MXFP4), RL (PPO, GRPO, SAC), interpretability (TransformerLens, nnsight).

GPU kernels from the metal up: PTX, tensor-core scheduling, cp.async pipelining. Plus mechanistic interpretability tooling (TransformerLens, nnsight) to trace what quantization quietly breaks.

newt & deuteron, a from-scratch Triton + Helion rebuild in ~4k lines of Python. Also poking at what quantization does to a model's calibration and factual recall, and an agentic-deception sandbox.

Collaborations on LLM efficiency and reliability, open-source GPU kernel work, paper discussions, and 1:1 mentoring. Always happy to talk quantization, inference systems, or compilers.