Blog · Topic
23 articles filed under infrastructure.
We publish the AI systems and operator knowledge we actually use—no information asymmetry between the desk and the community. Free blogs for doctrine; $1–$10 packs for checklists.
2026-08-11T00:00:00.000Z
The multi-lane training queue we use on free GPUs—claim, stage, checkpoint armor, pull home, prove, stop the VM. Privacy routing included.
2026-08-11T00:00:00.000Z
How Kaggle GPU/TPU quotas power DAJAI twin LoRA experiments—lane separation, privacy routing, and why free compute still requires operator rigor.
2026-08-11T00:00:00.000Z
How the DAJAI stack runs real workloads locally—Mac Studio brain, CyberPower GPU, tunnels, and why 'sovereign compute' beats rented GPUs for operator work.
2026-08-11T00:00:00.000Z
Everyone reaches for Postgres by reflex. I run a 299MB SQLite file as the analytics backbone for a whole creator operation — here's why one file beats a database server.
2026-07-16T00:00:00.000Z
Uploading thousands of files to an API is where naive scripts die. The resumable, self-healing pattern I use — idempotent state, OAuth refresh-token rotation, backoff, and a watchdog that doesn't lie.
2026-07-14T00:00:00.000Z
A local ML server that hardcodes its model breaks on every upgrade. Here is how I made mine model-agnostic — Qwen to DeepSeek-VL2 with zero code changes.
2026-07-13T00:00:00.000Z
How I stopped rebuilds from serving a broken site: atomic builds, a deploy lock, and manifest validation before cutover behind one Cloudflare tunnel.
2026-07-11T00:00:00.000Z
How I manage API keys across dozens of self-hosted daemons on one Mac — git-ignored dotenv with chmod 600, launchd wrappers, Keychain, gpg backups, no cloud vault.
2026-07-11T00:00:00.000Z
Self-hosting means owning your recovery. Here's the 3-2-1 backup rule, safe SQLite snapshots, encrypted offsite copies, and the restore drill everyone skips.
2026-07-11T00:00:00.000Z
Build a self-hosted uptime monitor that probes your own domains, checks SSL expiry, verifies launchd PIDs, and texts you only when something is actually down.
2026-07-11T00:00:00.000Z
No cloud host, no VPS. Three live domains served by one multi-tenant Next.js codebase on a Mac Studio, exposed through a single Cloudflare tunnel.
2026-07-10T00:00:00.000Z
One Mac orchestrates and runs vision; a cheap Windows GPU box does the heavy reasoning. The heterogeneous local-inference pattern — supervision, LAN wiring, and fallback.
2026-07-07T00:00:00.000Z
No Kubernetes, no Docker Swarm, no systemd. I run a production fleet of web servers, AI models, and monitors on one Mac using launchd — and it self-heals.
2026-07-05T00:00:00.000Z
Why a 3.19GB IQ3_XS quant that fits entirely on a 4GB GPU runs 2.5x faster than a bigger Q4_K_M that spills layers to CPU — a real local-LLM tuning log.
2026-07-04T00:00:00.000Z
I built my entire AI operation on local MLX models running on my own hardware and Cloudflare tunnels so I pay zero per token while keeping every prompt and every weight under my direct control.
2026-07-01T00:00:00.000Z
Stop renting a vector database. A self-hosted semantic-search stack on Apple Silicon: MiniLM embeddings, SQLite + cosine similarity, no API bill, no SaaS.
2026-06-19T00:00:00.000Z
A self-hosted wiki, glossary, and knowledge base belongs in your sovereign stack. The MLX + Next.js + Cloudflare-tunnel pattern, no SaaS, no rent.
2026-06-17T00:00:00.000Z
Three live Next.js sites, one Mac, zero VPS bill and zero open inbound ports. The self-hosted pattern: one named Cloudflare tunnel, launchd, and a dark origin.
2026-06-15T00:00:00.000Z
Three production websites, all served from a single Mac in Las Vegas, all reachable from the public internet through one Cloudflare tunnel. The pattern, the trade-offs, and why this is how sovereign infrastructure actually ships in 2026.
2026-05-10T00:00:00.000Z
A working sovereign AI stack on Apple Silicon in 2026 — the seven tools doing the load that used to live behind a credit card. What each one is for, what it replaced, and where the rough edges still are.
2026-05-10T00:00:00.000Z
After a year on the API meter, I moved the daily inference load onto an M4 Max running a 27B abliterated MLX model. The numbers, the trade-offs, and why sovereign AI is finally a serious answer in 2026.
2026-05-10T00:00:00.000Z
Hellcat Blondie on why your domain is the most underrated asset in the modern creator stack — and why every serious creator should buy theirs before someone else does.
2026-04-20T00:00:00.000Z