-
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.
-
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.
-
A passive GPU in a tower: VRAM hit 106 °C while the core throttled
My new 32 GB passive GPU had no plateau under LLM load and its memory reached 106 °C. Measurements, the limits that were silently ignored, and the fan, power cap and clock limit that fixed it.
-
ComfyUI on demand: systemd socket activation for a GPU container
Idle ComfyUI held up to 2.8 GB of VRAM my LLMs needed. A systemd socket now starts it on the first request and stops it after 10 idle minutes, and a guard makes room for image jobs.
-
Vision-model OCR that keeps the layout: what to trust it with
A vision model called a bold italic paragraph 'not bold, not italic'. My OCR now lets the model read, Tesseract place small text, and pixel measurement judge typography: 99% precision on italic words.
-
An AI studio that makes films and 3D prints in Blender
My AI agents build printable models and short films in a sandboxed Blender over a tool protocol. The hard part is that a text-only model cannot see its own output, so it ships broken geometry as 'watertight' unless you give it sight.