Software engineer with 8+ years in cloud infrastructure automation. Professionally I build platform automation and storage/backup orchestration; independently (and this is what most of my public work is about) I work on local LLM inference and fine-tuning.

Most of it runs on a home lab built for the purpose. I have a dual-RTX 3090 Debian machine I use to fine-tune, quantize, benchmark, and serve open-weight models. My main interest is inference efficiency and the ways we enable LLMs for the gradient of resources available to individuals and organizations. Quantization, speculative decoding, KV-cache and batch tuning, attention optimization, the works. I try to write the work up with enough configs and numbers to be reproducible rather than anecdotal.

• End-to-end fine-tuning - dataset construction → supervised fine-tuning (LLaMA-Factory, multi-GPU with DeepSpeed ZeRO) → RAG serving through Open WebUI, on a dataset I built myself.
• Inference optimization + benchmarking - comparative quantization studies (llama.cpp GGUF quants, AutoRound INT4, FP8 on vLLM), multi-backend serving via llama-swap, and KV-cache / batch tuning on consumer GPUs.
• Upstream bug reports - filed reproducible issues against llama.cpp and flash-attention from problems hit while running these models.

Detailed writeups - with configs, loss curves, and measured results are on my blog.

• Blog / writeups - https://demietrich.com/blog
• Hugging Face (dataset + models) - https://huggingface.co/dboybaker
• LinkedIn — https://www.linkedin.com/in/demietrich/

B.S. Computer Engineering, Washington State University (mathematics minor). Day-to-day languages: JavaScript, Python, Bash.