Sankur K.

I build and fix ML systems, and I check every result against a trusted reference before I hand it over. What I do • LLM training and inference: RL post-training (GRPO), quantization, serving, deterministic outputs (PyTorch, CUDA, C++) • GPU performance: tensor-core kernels, profiling, speed-ups verified against cuBLAS and PyTorch baselines • Data analysis in Python and Stata: cleaning, merging messy sources, panel data, reproducible results • ML tooling: evaluation harnesses, benchmarks, test suites Proof you can check • Nine fixes merged in 15 days into NVIDIA Ising-Decoding, Apple MLX, Meta ExecuTorch, CMU XGrammar, Hugging Face OpenEnv, Oxford PyZX and JaxMARL • A GPU compiler at 1.00–1.04× cuBLAS speed, and an LLM engine that matches vLLM token for token • Integer-only LLM inference that gives identical bits on an NVIDIA GPU and a CPU • On Upwork: rated 5.0 by a research team in Spain. I merged two hotel censuses (5,909 and 8,058 records) and built a 10,038-city-month Google Trends panel. How I work I state my assumptions up front, give you tests or reproducible scripts with every delivery, and send short progress updates. I’m in India (IST) and can overlap with European mornings and US evenings. Good fit: LLM or GPU performance work, ML pipelines, research data cleaning and analysis. Not a fit: generic data entry, or projects with no clear success measure.