Aria — Local Offline Personal AI Assistant

The Problem

I wanted a personal assistant that never sends data to cloud LLMs and can run real-time voice, vision and memory features on an ordinary home PC with just 8GB of VRAM.

The Approach

Wrote a router that picks between 5 different local models by task type (fast/code/deep-reasoning/vision/uncensored mode), with a mechanism that unloads unused models from memory to respect the VRAM budget. Used ChromaDB to embed conversation history for RAG-based memory.

Engineering Highlights

• Task-based automatic routing across 5 chat models plus an embedding model.
• ChromaDB-based RAG memory with deterministic SHA-256 IDs for consistent records.
• Fully offline voice interaction via Whisper (local STT) + Piper TTS.
• A two-tier regex security filter blocking dangerous commands (rm -rf, powershell -enc, etc.) — hardened after closing a documented vulnerability.