Argus, my local agent system, and a llama.cpp latency fix
- A personal assistant on a local 35B mixture-of-experts model (Qwen3.6-35B-A3B, llama.cpp on Vulkan, AMD integrated GPU) inside the open-source Odysseus hub. I wrote the MCP memory server (seven tools, confined paths, no deletes), the voice pipeline, a supervisor that restarts failed services, and a learned router that sends hard jobs to Claude Code and keeps private data local.
- A weekly report card on hidden test items, with confidence intervals, compares the bare model, the assistant and the assistant with deliberation. A new prompt habit is adopted only after it also wins on a separate held-out set.
- Settings chosen by evaluation on 13 everyday tasks with deterministic checks. Turning the model's thinking off scored 13/13 at a 2.6 s median, against 10/13 at 19.8 s with it on.
- Traced a 3-second pause before each spoken answer to the Vulkan backend reading uncached GPU memory. A one-line condition cut a state snapshot from 0.61 s to 0.03 s, bringing the first spoken word to about 0.7 s with identical output. The same fix had reached upstream two weeks earlier; I found it independently.
sergioadan.dev/argus · sergioadan.dev/work/llama-cpp-stall
How it was measured
A test server on a spare port with production flags, verbose timestamped logs, then temporary timers on the read path; the fix was verified with a scripted six-turn conversation at temperature 0 against the stock build, word for word.