Index
Writing
RSS →Deep dives with real numbers — medians and ranges, the line between what I observed and what I inferred drawn out loud. Fewer, longer, sourced.
The anatomy of time-to-first-audio in a voice agent
From turn-commit to first audio: LLM ~59%, TTS ~28%, my orchestration ~14% — and one first token took 2.9 s. I was tuning the wrong layer.
Giving AI agents a real phone line
50 concurrent phone calls in userspace Go: p50 held at 20.00 ms at every load level — measured, not claimed. The tail is where it gets interesting.