AI that runs in your browser
Ask anything about Yuta Hidaka. A small AI model running inside your browser answers. The model loads once and is then cached in your browser — no re-download afterward. You can also pick a no-download “No LLM” mode for instant answers.
Loading the model into your browser…
Mind your data usage (Wi-Fi recommended). It downloads once and is cached, so it is fast next time.
No server, no API key. Everything runs inside your browser.
Model (selectable)
Qwen2.5 / TinySwallow 1.5B
Defaults to Qwen2.5-1.5B (~0.9 GB) — the strongest English answers here; Sakana AI's TAID-distilled TinySwallow-1.5B covers Japanese. Both target WebGPU-capable desktop browsers; phones (incl. iPhone Safari) automatically fall back to a low-memory Qwen2.5-0.5B GGUF running on llama.cpp (~0.4 GB). The model loads once, then is cached; a no-download “No LLM” mode is also available.
Runtime
transformers.js / WebLLM · WebGPU
Hugging Face · ONNX / MLC
The weights are 2–4-bit quantized — that is how a multi-GB model fits in ~0.5–0.9 GB, and why further compression does not help. The 1.5B models run on WebLLM (MLC, WebGPU-only, the same stack as Sakana's official demo); the light builds run on transformers.js (ONNX) and wllama (llama.cpp WASM). Weights stream from Hugging Face once, then stay cached in your browser. Everything runs in a Web Worker, so the UI never freezes.
Grounding
Retrieval + fact-check
RAG + verify
Only the résumé facts relevant to your question are retrieved as context. After generation, names, numbers, and credentials are verified; anything unconfirmed is dropped and replaced with the real data.
Static site (Astro) · nothing sent to a server · no API key