Private Intelligence. Local First.
Everything runs on your hardware. Nothing leaves your machine.
Wake-word detection, Silero VAD, and faster-Whisper for speech-to-text. Zero-shot voice cloning via OmniVoice TTS — speaks in a natural voice with sub-200ms latency.
STT + TTSContinuous screenshot and camera analysis using local vision models (Gemma 4, Qwen3-VL). Proactive commentary and context-aware responses based on what's on screen.
Multi-modalAmina schedules her own wake-ups and sends reminders via Telegram. Full task management with /tasks, /snooze, /reschedule, /done, and /history — no server required.
TelegramGenerates stories chapter-by-chapter with matching images. Produces music via ComfyUI, renders video with LTX, and composes multi-track playlists with DJ intros.
Image · Video · Music · StoryTri-store memory system: short-term ring buffer, episodic journal with semantic search (sentence-transformers), and a structured profile of long-term facts per user.
Episodic + SemanticAll inference runs locally via llama.cpp. User data is isolated per account with optional 7-Zip AES-256 encryption at rest. No API keys, no telemetry, no cloud.
Local · EncryptedSeparate memory, personality, and media per user. GPU resource ownership with conflict resolution — competing requests queue intelligently without crashing services.
Session-IsolatedA pi.dev terminal bridge with an agentic coding assistant. Plan, execute, and iterate on tasks in a sandboxed environment — with optional desktop control mode.
AgenticAudio, images, and video in this demo — all generated by Amina.
Amina made it. Local inference, no APIs, no cloud.
All code is publicly available. Self-host Amina on any machine — capabilities scale with the VRAM you throw at it.
★ View on GitHub 📄 Setup GuideBuilt on open-source inference — no proprietary cloud services.
This system is closed to the public. Account creation is disabled — only authorized administrators may proceed.
Administrator Login