~/dev/raymond$loading
Every developer-portfolio uses-page reads like a wishlist. This one is the actual machine. If you see something here, it’s because I’ve sat in front of it and shipped something with it — not because the marketing site said it was good.
The whole reason this site mentions “local-first AI” without laughing. The 24 GB of VRAM is enough to run Gemma-class 26B quants alongside ChromaDB, Whisper, and a Stable-Diffusion-shaped workload at the same time. Most of my projects assume this GPU is sitting there.
On-demand, mostly for working on the deck or out of the house. Talks to the desktop over a Tailscale mesh — Syncthing keeps ~/developer/dev in sync with C:\Dev, and rclone-over-SSH mounts the Windows disk at ~/MountedDev for the bigger files Syncthing skips.
Every box I own is on the same tailnet. The Windows file server runs rclone serve sftp on a Tailscale-only address; nothing is exposed to the public internet. Best “it just works” software I’ve adopted in years.
Editor is VS Code. The agentic layer on top is Claude Code. It also drives every LLM call inside Job-Shorts via a subprocess — that's how I get 42 chapters rendered without paying per-token fees.
PowerShell for native Windows things, bash for everything POSIX-shaped (Git Bash on Windows, real bash on the Mac). Both running through Windows Terminal with a JetBrains Mono / Fira Code stack.
uv for Python env / package management — fast enough that it stopped being something I think about. rufffor lint + format because the <100 ms feedback loop changes the shape of how I write Python. ESLint + Prettier on the TypeScript side.
Code on github.com/Raymondriter. Sites on Vercel (this one included). Six Python packages on PyPI.
Default endpoint for every project that needs an LLM. Currently running a Gemma-4 26B quant most days. The OpenAI-compatible API means every project can point at http://localhost:1234/v1 and not care that it's local.
Vector store for GBC-AI, LTVRAG, and a couple of other RAG projects. Persistent on disk, fast enough on a single thread that I've never needed to scale it out.
STT for both GBC-AI (offline sermon transcription) and LTVRAG (real-time voice). The turbo variant is fast enough to feel responsive without the accuracy fall-off of the smaller models.
Image and video gen for Job-Shorts. ComfyUI for the graph execution, LTX 2.3 for actual video synthesis, F5-TTS for narration. All local — no SaaS in the loop.
Vision stack for rsbot and The Visual Bridge. YOLO trained on my own gameplay screenshots, Moondream2 as the “describe this scene” fallback when YOLO is unsure.
App Router everywhere new. MDX for blog posts. Tailwind for styling. next/og for the terminal-themed OG cards you see when you share these pages.
GBC-AI, LTVRAG, sportbetting, ai-world — all FastAPI. Async, typed, fast, and never gets in my way.
Supabase for new projects (Postgres + Auth + Realtime + Edge Functions). Firebase for legacy projects already on it. helpmetopray.org is the cleanest example of the new pattern.
Doesn't technically run code, but it's on the rig list because the peppered beef jerky it makes on a 4-hour 180° smoke is genuinely better than anything I can buy. Eye of the round, hand-sliced, Dragon's Milk stout in the marinade.
The reason GBC-AI exists. Years of sermon archive, finally queryable.
My wife steers half my project list. My son Jace has his own CRM because he wanted free swag. They are the actual reason most of this exists.