Experiments
Local AI lab
Tests and prototypes, mostly with models running locally: edge inference, agents, offline RAG.
Experiment · Inference
Less waiting, more agents: upgrading my local AI stack
From Lemonade and Vulkan to llama-swap, Docker and Halogen: how I upgraded my local AI stack to put agents to work. Performance, decisions and results without the alphabet soup.
Sep 2026
Experiment · Inference
Running Qwen3.8 27B locally: when the model works but the chat cuts out
How we configured Qwen3.8-27B in HalostrixLab and why updating Lemonade fixed a streaming disconnect. Local models, context and SSE explained from the ground up.
Sep 2026
Experiment · Inference
Vulkan or ROCm on Strix Halo: how I failed, measured, and chose a local inference backend
A llama.cpp experiment on AMD Strix Halo: UMA memory, failed smoke tests, reproducible gates, and a short Vulkan versus HIP/ROCm comparison.
Sep 2026
Experiment · Training
From a ROCm smoke test to a fine-tuning lab: what Docker proves—and what it does not
How I am building a local AMD training lab: one real backward pass, one configured workspace, one partly validated integration, and explicit limits.
Sep 2026