A live update
on work and life.

I've been building out a small mesh of home servers to run local AI models and side projects, and lately the interesting work has been in two threads: a D&D campaign-management app called TheLongRest, and the plumbing that keeps a fleet of self-hosted models actually reliable.

On TheLongRest, most of the recent work has been about what an AI assistant is and isn't allowed to say. It answers rules questions and generates session recaps, but D&D content splits into what I can legally redistribute (the SRD, under CC-BY) and what a player might paste in from a copyrighted rulebook. I built a note-taxonomy system so a DM can mark individual paragraphs in their session notes as player-visible, DM-only, or "reveal later" — and the retrieval layer respects that at the chunk level, not just the document level, so editing one paragraph can't accidentally leak a held-back plot twist to the whole table. Along the way I found and fixed a real leak: re-indexing a document was silently dropping the visibility tag, so an innocuous edit to a DM-only page made it queryable by players. Absence of a tag was being read as "safe to show," which is exactly backwards for anything touching secrets.

I've also been chasing down a case where the AI assistant just made something up — it described a paladin subclass feature that isn't in the SRD, invented a plausible-sounding source book for it, and got it wrong. Tracing that back took a few layers: the system prompt was overclaiming what books it actually had access to, the fallback behavior let it guess instead of admitting a gap, and separately, the character importer had been silently dropping subclass features on import, so the AI didn't even know the character had that ability. Fixing all three closed the loop — it now says "I don't have the text for that, but I can see your character has it," which is a more honest answer than a wrong one.

On the infrastructure side, I run a couple of GPUs on repurposed desktop hardware serving local language models for chat, image generation, and voice, all mediated through a small proxy I wrote that queues and hands off GPU access between whichever service needs it. The card is old enough (a Tesla P100) that plenty of assumptions from newer hardware don't hold — expected throughput numbers from bandwidth specs are just wrong on this generation, and I've had to actually benchmark instead of extrapolate. I also learned the hard way that Docker's "restart unless stopped" policy treats a process that was stopped for any reason, including one it stopped itself, as intentionally down forever — so a service can sit dead through a reboot with nothing logging an error. That one cost an hour of silent downtime before I found it.