Local inference
Qwen3.5-35B-A3B stays warm on a single RTX 3090 through a loopback-only llama.cpp service.
Private AI · local-first PWA
Pocket Search
Search the web, reason with a local Qwen model, and keep the resulting conversations on your own workstation.
The working app is intentionally tailnet-only. This public page does not proxy chats, prompts, or search traffic.
Inspectable research
SearXNG retrieval, hostile-page isolation, persisted evidence, and validated numeric citations.
Durable mobile UX
Installable Android shell, offline drafts and history, reconnect-safe streaming, retry, and cancellation.
Private boundary
Signed same-origin sessions, strict CSRF checks, loopback dependencies, and Tailscale Serve HTTPS.
Production shape
One workstation. One private URL.
React and FastAPI form one same-origin application. SQLite makes turns durable before generation starts; SearXNG and llama.cpp remain invisible to the browser. A service worker keeps the shell useful when the phone disconnects without ever caching private API responses.