Pillar: Self-Hosted LLM Infrastructure

Downloading ComfyUI Models Without the Pain Hugging Face has a model download UX designed for interactive notebook sessions, not production pipelines. If you’ve ever watched a 12 GB checkpoint stall at 94% with a 403 error, waited for the browser download bar to reach 8 GB before realising the connection dropped, or tried to resume […]

RTX 3090 ComfyUI Tuning: xformers + sage-attention Default PyTorch attention is safe and predictable. It is not fast. On the RTX 3090 with SDXL, switching from vanilla attention to xformers cut per-image generation time by roughly 25%. Adding sage-attention on top trimmed another 10–15% on longer diffusion runs. Neither change requires touching a single ComfyUI […]

ComfyUI on an RTX 3090: a Local-First Setup Running image generation locally isn’t about avoiding API costs — it’s about having a workflow that doesn’t change under you. Cloud services deprecate models, change pricing, and throttle batch jobs. A local ComfyUI setup on an RTX 3090 stays the same until you decide to change it. […]

Running Claude Code with GLM-4.6 (Ubuntu + Codespaces) Claude Code doesn’t have to run against the Anthropic API. It supports any OpenAI-compatible endpoint — which means you can point it at a locally hosted model. GLM-4.6 from Zhipu AI is a capable open-weight model that speaks the OpenAI chat format, and getting it working with […]

Telegram/WhatsApp gateway into local models.

A 3-node self-hosted AI platform — local LLM inference, a GPU media/voice node, and a containerized hosting estate — that builds and runs real businesses at near-zero inference cost.