Say It, Don't Type It: Kadence, a Voice-Only TPM for Linear

Building a self-hosted, sub-two-second voice agent for Linear — Whisper, Qwen 2.5 and Kokoro on one GPU, with constrained decoding that makes a hallucinated action structurally impossible. What worked, what broke in real meetings, and why I stopped.

Prefill, Decode, and the Quiet Third Job: Reviewing vLLM's Guide to Disaggregated Serving

Why long prompts make every stream stutter, when splitting prefill from decode pays off, and why tokenization belongs off the GPU, with a first-principles simulator that reproduces the guide's benchmark.