Local Jarvis AI Assistant — RAG + Local LLM + Vision (Personal)
Designed and deployed a fully offline voice-activated AI assistant using a local LLaMA model with a retrieval-augmented generation (RAG) pipeline. The system maps to LLM security training concepts such as prompt injection and tool-call abuse vectors by structuring how inputs and retrieved context influence responses. While no explicit supervised fine-tuning dataset is described, the project operationalizes response generation from labeled/grounded retrieval content. • Implemented a RAG pipeline with local vector embeddings for context-grounded responses • Integrated Vosk STT, OpenCV for vision, and text-to-speech for multimodal interactions • Structured the assistant architecture to test/handle prompt injection and agent-based SSRF patterns • Produced voice-command-to-response behavior using local model inference