-
Why I went fully local: from cloud AI to a sovereign lab at home
In April 2026 my AI setup mixed slow local models with cloud ones. By October everything that runs, stores or decides was local. The library-scale arithmetic, the test that ended the argument, and the one exception I still keep.
-
Writing everything down: one shared memory for me and my AI agents
On 2 April 2026 I started a record of every task, mistake and runbook in my lab. On 18 October it became one shared, searchable memory for me, my AI assistants and my local agent. What it meant to work with AI in the open record.
-
Four agents, four lessons: how my local AI agent grew up
I built four AI agents on one home server in seven months. Each failed in a new way: code that was not running, reports that said all clear, tasks closed with no work, a deleted home folder. Each failure moved trust out of the model.
-
Letting my AI agent act: autonomy with guardrails, a work in progress
After an AI test wiped my home folder, my agent could only propose. On 11 October I let it act alone on one reversible action, behind a deny list, an allow list, dry runs and a typed second opinion. The design, the first run, and my own mistake.
-
13 hours of question marks: a canary for LLMs that look healthy
After a llama.cpp router restart my main model answered every request with question marks for 13 hours while every health check stayed green. A known-answer canary now catches it within 10 minutes.
-
How small local models finish hour-long agent tasks
Left alone, my 9B agent closed tasks without doing them and invented a 10 KB report from an empty sub-agent result. Scaffolded task trees, evidence checks and deterministic gathering let it finish 66-minute projects.
-
One GPU, many LLMs: the VRAM guards that stopped a driver deadlock
With 33.9 of 34.2 GB of VRAM in use, a desktop app triggered a 24.7 GB eviction and froze my GPU driver. Here is the VRAM budget and the two guards that now enforce it.
-
Agent skills as SKILL.md playbooks, not plugins
My agent's plugin folder auto-loaded code as ungated tools. I moved procedures to SKILL.md playbooks in one git library: validated in CI, root-owned, read-only to the agent, and promoted only after review.
-
Local AI study guides for thousands of books, capped on purpose
Hosted study guides cost 80k tokens per part, so I moved to a local 4B model. What broke, the guardrails that fixed it, the 100-chunk and 60-section caps, and why breadth beat depth.
-
From Ollama to a llama.cpp router: 49 models down to a few presets
I replaced Ollama with a llama.cpp router on my home-lab GPU: on-demand presets, sleep timers, embedding parity tests, and a log-driven retirement that freed about 400 GB.
-
Swapping an 8 GB NVIDIA card for a 32 GB AMD GPU: what broke
I replaced an RTX 4060 Ti with a 32 GB Radeon AI PRO R9700 for local AI. Ten things broke, mostly silently; here they are with fixes, and the before-and-after throughput.
-
Hybrid search with bge-m3 and Elasticsearch: six lessons
What I learned building BM25 + kNN hybrid search with reciprocal rank fusion and bge-m3 on one shared GPU: an analyser mismatch (38 to 2,631 hits), RRF pitfalls, embedding timeouts and a highlighting limit.