Everywhere else on this site, if a demo is shown, it's real and computes what it claims to. This page is the opposite: things genuinely not built, stated as such, with no preview mockup underneath pretending otherwise.
Running a quantized small LLM and ONNX voice synthesis entirely client-side, via WebAssembly + WebGPU, with no server round-trip at all.
This is a real engineering project on its own (model quantization for browser targets, WebGPU compute shader work, a completely separate inference path from the production Ollama/Rust pipeline) — not a small addition to the existing system. Nothing here has been built or benchmarked yet.
A one-click hosted notebook flow for voice fine-tuning and benchmark sweeps on free cloud GPUs, beyond the existing manual Colab notebooks already in `notebooks/`.
The manual notebooks work today and are the current supported path (see the Voice section above). Automating the click-through flow is a nice-to-have, not yet scoped or started.
A shared, opt-in registry where people could publish and browse authored persona presets beyond the built-in examples on the Showcase page.
This is a local-first, single-friend-per-person project by design (see the community product decisions in the engineering ledger) — a public registry raises real moderation, consent, and hosting questions that haven't been worked through yet, not just an engineering task.
Short, real GPT-SoVITS-rendered audio clips for the Voice Showcase page, replacing the current static parameter table.
Generating these requires running the actual voice pipeline against a consented voice sample and exporting clips as static assets — a content-production step, not just code. The parameter table shown today is accurate; it just doesn't play audio yet.