Posts
Reviews of AI tools, agents and self-hosted software I have run on real work, the monthly token market, and notes from the homelab. Newest first.
Showing all posts.
-
Aider with local models through Ollama: two weeks on my own backlog
Worth a trial.A 32B local model handles small, well-described edits; for refactors across files I still reach for a hosted model.
-
AI token market, November 2026
Takeaway:Output tokens still cost several times input across the hosted tiers; hosted open-weight models stay the cheap floor.
-
Backups I actually test: restic, a second disk and a monthly restore drill
Keep using.Boring, cheap and it restored everything I asked of it. The drill matters more than the tool.
-
Letting a coding agent open its own pull requests for a week
Skip for now.Review time ate the time it saved. Useful for dependency bumps, not for features.
-
AI token market, October 2026
Takeaway:Cached-input discounts now matter more than headline prices for agent workloads that resend the same context.
-
Whisper on a CPU-only box: transcribing meetings overnight
Keep using.Good enough for notes by morning. Far too slow for live captions without a GPU.
-
From Docker Compose to single-node k3s, and back again
Skip for now.Eleven services on one machine did not need a scheduler. Compose and a systemd timer were enough.
-
AI token market, September 2026
Takeaway:Batch pricing is the quiet bargain: if a job can wait a few hours, it usually costs about half.
-
A self-hosted chat front end for a five-person team
Worth a trial.Single sign-on and per-user spend limits worked; check the licence terms before you rebrand it.
Sample post list. Only the Aider review is written out in full on this build.