ThinkSpread
ThinkSpread is where I publish independent local-AI work.
Apple Silicon Macs are no longer small machines. They have fast GPUs, a Neural Engine, and enough unified memory to run serious local AI workloads.
But most local inference runtimes are still shaped around one user, one prompt, one stream. They lack the server-side machinery agents need: continuous batching, shared KV, SSD KV offload, and scheduling for many concurrent requests. That matters when several agents repeat the same system prompt and tool history: the runtime should not pay for the same context over and over. Vision-language models leave another gap: the Neural Engine is often idle while the GPU does all the work.
Tempo9
The first project is Tempo9: server-style inference for local AI on a Mac you own. Continuous batching, shared KV, SSD KV offload, multi-request scheduling, and GPU + Neural Engine execution, so multiple agents can share more work and duplicate less memory behind an OpenAI / Anthropic / Ollama-compatible endpoint.
Voice Rack · 声音档案馆
Ask your recordings. Offline. A private archive for everything you have ever recorded — transcription, cross-file search, and answers with clickable timestamps, entirely on your machine. One-time purchase, no subscription.
Overview · In development · built on Tempo9
MirrorMirror
Point a Mac's camera at the world and a local vision-language model narrates it, one frame a second — vision on the Neural Engine, language on the GPU, no cloud anywhere.
Overview · working demo · built on Tempo9
Writing
Chinese and English notes on measurements, false starts, and the hypotheses the numbers killed. What did not work is published alongside what did.