• Wireless pairing

  Scan a QR code, the phone downloads and installs the app via browser, opens it, scans a second pairing QR, and the phone reports its own real hardware specs (RAM, CPU arch, storage, Android version) to register itself. No USB, no ADB at any point.

• Natural-language model selection

  An AI assistant ("Loki") takes a request like "I want real-time object detection" or "give me an offline chatbot", searches HuggingFace for real candidate models, filters them down to what actually fits the paired phone's RAM budget, and explains the top picks in plain English — it will honestly say "nothing fits" rather than hallucinate a model that doesn't exist.

• One-click wireless deployment

  Pick a model, hit Deploy, and the backend pushes a download instruction over WebSocket; the phone pulls the model file directly from the HuggingFace CDN itself. The backend never stores or touches the model file — only URLs.

• Three on-device inference modes

  Camera (TFLite — live object detection/classification with on-screen labels), Chat (GGUF via llama.cpp — real token-by-token generation), Voice (TFLite audio — mic-based sound/speech classification).

• Offline after deployment

  Once a supported model is on the device, the inference path is designed to run without internet access, verified in airplane mode during development. Activated inference does not require a remote model API per request.

• Fleet dashboard

  See all your paired devices, their live specs and connection status, deployment history, and a chat panel for the assistant, all backed by real data (no mock/demo content anywhere).

• Production-grade accounts

  Email/password signup with OTP verification, forgot/reset password, and hardened JWT auth with short-lived access tokens, httpOnly refresh cookies, rate limiting, and locked-down CORS.

LokiAI does not intend to charge per inference. Customers still carry hardware, power, storage, maintenance, and operational costs.