Authors: Hitesh Vinod Dadlani, Ayaan Badshah Khan, Munaf Irfan Shaikh, Professor Jayshri Kandekar
Abstract: The rapid advancement of Large Language Models (LLMs) has significantly expanded the capa-bilities of intelligent software agents. Traditional virtual assistants are limited to reactive query-response interactions and lack persistent memory, flexible tool integration, and autonomous task execution. This paper presents AVIA (Autonomous Virtual Intelligent Assistant), a locally-hosted multi-agent multimodal assistant designed for autonomous digital task automation through an extensible skill-based architecture. AVIA integrates LLM reasoning with a persistent mark-down-based memory system, a modular skill execution framework, and proactive scheduling via heartbeat loops and cron-based task management. The system supports multimodal interaction through text and voice interfaces and integrates with external platforms including email, calendar, document management, and social media. A local-first design philosophy ensures user privacy and independence from cloud infrastructure. Experimental evaluation across five task cate-gories demonstrates a Task Completion Rate (TCR) of 91.3%, an Automation Success Rate (ASR) of 88.6%, mean response latency of 2.8 s, and memory retrieval accuracy exceeding 93%, validating AVIA as a flexible, privacy-preserving foundation for next-generation intelligent personal assis-tants.
