Convo-AI

I wanted a conversational AI assistant that worked entirely locally — no cloud API keys, no telemetry, no data leaving my machine. It needed to listen to voice input, think with a local LLM, remember context across conversations with RAG, and speak responses back out loud. The kind of thing you could run on your own hardware and trust.

Convo-AI is a local-first, voice-enabled conversational AI assistant with a full pipeline: it listens with faster-whisper, thinks with a local Ollama LLM, remembers with RAG over SQLite, and speaks with Coqui TTS (XTTS v2). No cloud API keys and no telemetry. The server is a FastAPI WebSocket backend in Python, with a Python CLI client and a React + Tailwind web UI. Configuration lives in config.json for voice model, speed and pitch, WebSocket URL, and LLM settings.

• Solo developer — built the entire voice pipeline and infrastructure
• Integrated faster-whisper for speech-to-text
• Connected Ollama as the local LLM backend
• Implemented RAG memory over SQLite for conversation context
• Integrated Coqui TTS (XTTS v2) for text-to-speech output
• Built the FastAPI WebSocket server and Python CLI client
• Built the React + Tailwind web UI
• Added Docker support for containerized deployment

• GitHub repo: github.com/BradleyMatera/Convo-Ai
• Live demo: bradleymatera.github.io/Convo-Ai
• Created: April 16, 2025
• Commits: 15 commits
• License: MIT
• Requirements: Python 3.8+, FFmpeg, Ollama, microphone for voice input
• Usage: python server.py (start server), python talk.py (run client), 'r' to record
• Config: config.json for voice model, speed & pitch, WebSocket URL, LLM settings
• Honest limitations: Requires local installation of Ollama and FFmpeg. Requires microphone for voice input. No cloud deployment — local-only setup. Small project (15 commits).