Blog
Got a topic you’d like me to cover? Feel free to reach out!
Three and a Half Weeks, 689M Tokens, No Meter
689M tokens. 25 days. ~$40 electricity. The math, the caveats, and what changed when the meter stopped existing.
Qwen3.8 27B on Dual RTX 3090s: A vLLM Field Report
Three days into making Qwen3.8 27B my daily local coding driver on dual 3090s — the hybrid linear-attention architecture, an OOM that masqueraded as a backend error, and what the …
I Built an LLM Traffic Router for My Homelab — Here is What Happened
I hooked up multiple local models behind a single proxy that routes traffic by complexity. Here is what I found deploying it on a cheap Intel GPU.
clean-term: A Minimal Oh My Zsh Theme That Actually Works Everywhere
Every terminal theme I tried either broke across fonts or added visual noise I did not need. So I built a 14-line ASCII-safe Oh My Zsh theme that works anywhere.
Locking Down a DMZ Web Server with Cloudflare Tunnels
Hardening an internet-facing web server sitting in a home DMZ: firewall, SSH, SELinux, fail2ban, and tunnel-only traffic via Cloudflare.
How I Bridged CLI AI Agents to Apple's Walled Garden
Apple apps are locked down with no APIs, but SQLite local databases and AppleScript let CLI AI agents bridge Reminders, Notes, Calendar, Mail and more — no cloud, no API keys, just …
Why I Use Both GitHub and Forgejo - And Why That Matters
I use GitHub for public collaboration and Forgejo for self-hosted control. They share the same Actions workflow syntax, just in different folders. Here is why running both makes …
Self-Hosting Plausible Analytics with Podman and Cloudflare Tunnels
How I deployed self-hosted Plausible analytics on a podman server, routed it through Cloudflare tunnels, and integrated it with my Hugo site — with a backup strategy that actually …
Self Hosted AI: Actually Running Local LLMs for a Multi-User Household
How self-hosted AI became the final piece of my homelab puzzle, delivering true parallel processing for multi-user setups and unlocking the real superpower of knowledge management.
Running Qwen3.6 27B Locally on Dual RTX 3090s with vLLM v0.19
How I went from a blank Docker template to 116+ tok/s with speculative decoding, FlashInfer, and a 160k context window on dual 3090s.
Showing 1-10 of 63 posts (Page 1 of 7)