# Ramshankar Bhuvaneswaran

> AI Engineer. I ship LLM systems with customers: agent workflows, RAG over real ops data, and the eval harness that keeps them honest in production. Also an ML systems engineer: training and inference infrastructure. Portfolio and technical writing.

- Site: https://ramshankar07.github.io/portfoliov3/
- Contact: bhuvaneshwaran.r@northeastern.edu
- LinkedIn: https://linkedin.com/in/ramshankarb
- GitHub: https://github.com/Ramshankar07
- Schedule: https://cal.com/ramshankar07

## Focus

- Training stack: PyTorch, FSDP, verl/GRPO, distributed orchestration on Ray, Kubernetes, SkyPilot, SLURM
- Inference stack: vLLM, ONNX, CUDA/HIP/Triton kernels, CuteDSL and PTX on Blackwell NVFP4, ROCm on AMD MI300X
- Benchmarking and profiling: timing-and-correctness harnesses, roofline/bandwidth measurement, regression evaluation
- Upstream: HuggingFace transformers PR #46084 (merged), LLVM libc PR #175396 (merged), vllm-project/llm-compressor PR #2323 (open)
- Research: BitSkip, https://arxiv.org/abs/2510.23766
- Alumni of Northeastern University (MSIS) and Anna University (B.Tech CSE)

## Key pages

- Home: https://ramshankar07.github.io/portfoliov3/
- Blog — Llama 3.1 budget fine-tuning: https://ramshankar07.github.io/portfoliov3/blogs/llama3-budget-finetuning.html
- Sitemap: https://ramshankar07.github.io/portfoliov3/sitemap.xml