Aster
A fully local, privacy-first agentic runtime powered by a single llama.cpp / llama-server instance — ~68 tools, a 15-round autonomous tool-calling loop, and strict hallucination guardrails serving CLI, Telegram, Discord, and WebRTC from one model load.
- Architected and built a fully local, privacy-first agentic loop featuring ~68 tools, a 15-round autonomous tool-calling loop, and strict hallucination guardrails serving CLI, Telegram, Discord, and WebRTC protocols from a single model instance.
- Engineered on-device audio input/output pipelines integrating Faster-Whisper for low-latency speech recognition, Kokoro TTS for voice synthesis, and real-time voice barge-in handling via LiveKit Agents.
- Integrated local multimodal perception via YOLOv8 and OCR screen/webcam inspection alongside a 5-signal emotion fusion engine to adapt assistant responses dynamically.
- Implemented persistent agentic memory using ChromaDB vector search and markdown contextual indexing running fully locally on consumer GPU hardware.
~68 tools · 15-round loop · 5-signal emotion fusion · one llama-server instance