Blog

My Obsidian-based writing setup was publishing beautifully, except synced posts landed in Typecho full of raw HTML clutter — a hidden XML-RPC setting fixed it.

A CSS/plugin conflict was rendering code blocks as an ugly blank space — turned out to be a syntax-highlighting plugin fighting the theme.

Running ten locally-hosted models on consumer GPU hardware to see which ones best summarize customer-service calls after speech-to-text.

Installing curl via snap put it ahead of /usr/bin on PATH, which broke the Ollama installer in a confusing way.

After one too many driver crashes from installing random open-source software, I wiped the machine and set it up as a single-purpose LLM box.

Working through two ML fundamentals in plain language: what a probability density function actually tells you, and how weighting by probability density changes mean squared error.

What an ‘agent’ actually means in the context of large language models, and a worked example in a customer-service setting.

Notes on getting an NVIDIA driver and CUDA 12.4 installed cleanly, including disabling the nouveau driver first.

A one-line fix for a CUBLAS_STATUS_NOT_INITIALIZED crash: just update Ollama.

After AMD’s ROCm ecosystem let me down for local fine-tuning, free Colab GPUs turned out to be the pragmatic way to fine-tune Llama 3 for free.

Self-hosting the open-source MaxKB knowledge-base tool at home, and pointing it at a Qwen-backed model for Chinese-language documents.

Pairing a consumer GPU running local Llama 3 with Open WebUI and One API to combine local and remote models behind one interface.