About
I'm Rodney Woods, an AI engineer in Victoria, BC. For the last two years I've been building a fully offline speech and language platform (ASR, TTS, voice cloning, translation) for a low-resource language: data pipeline through model training, evaluation, and deployment, solo, on a single consumer GPU. Before that I designed, built, and operated a 250 kW GPU compute facility, also solo.
The through-line of this blog: generative models are useful exactly to the degree you can measure them. I gate every generated output with measurements calibrated against ground truth, because I work in a language I cannot proofread by ear. Most of what I write here is what that discipline catches.
Lately that discipline has pointed at inference itself: quantization and serving on consumer Blackwell, where the gap between what a benchmark says and what actually ran turns out to be measurable. Bugs found along the way get filed upstream with fixes offered (llm-compressor #3011, flashinfer #4410).
Selected work, all reproducible:
- sm120-quant-bench: NVFP4 and W4A16 quantization on an RTX 5090, with capability probes that catch the damage aggregate metrics hide and a kernel receipt attached to every number. Write-up: NVFP4 on a 5090 names two different machines.
- local-tts-router: multi-model local TTS routing with automated round-trip CER quality gating, best-of-N against a Whisper judge.
- voicegate: deterministic style-conformance gating for prose, calibrated from an author's own writing, with measured deltas instead of vibes.
- throughline: deterministic continuity checking for fiction manuscripts, timeline and consistency bugs in milliseconds, no LLM in the loop.
- KinSpeak: the offline speech platform above (ASR, TTS, verified translation); code and models stay with the community it serves, so the site is the public surface.