SAARTHI: Fine-Tuning A Sub-Billion-Parameter Language Model for Offline Voice-Controlled Home Automation Using Lora and OLLAMA

Article Citation : Jalaj Dabas , Hardik Arya , Mandeep Sindhu , Kapil Sharma , (2026) "SAARTHI: Fine-Tuning A Sub-Billion-Parameter Language Model for Offline Voice-Controlled Home Automation Using Lora and OLLAMA", International Journal of Industrial Electronics and Electrical Engineering (IJIEEE) , pp. 1-8, Volume-14,Issue-7

Abstract : All popular voice assistants, from Alexa, Google Hometo Siri, require sending audio to are mote server first before any action can be taken. The use of remote infrastructure creates additional latency, fails in case of internet connection disruptions, and sends voice data to third-party servers. Saarthiis an offline home automation assistant that operates solely on the user‘s local machine.ItispoweredbyQwen2.5-0.5B-Instructfine-tunedusing LoRA on 1169 English voice commands spanning binary device control and continuous level-based actions across five household appliances. After integrating the adapter and quantizing the weightsto4-bitGGUFformat(Q4KM),the model size of 320MB processes input via Ollama on CPU in just 1–3 seconds. An ESP32 microcontroller executes actions with relays and IR signals. All responses are filtered through a type-aware validation layer. In a 50-command held-out test, the model achieved 94% accuracy with no false actuations. Keywords – Large Language Models, Low-Rank Adaptation, Fine-Tuning, Home Automation, Ollama, Quantization, Edge AI, Speech Recognition

Type : Research paper

Published : Volume-14,Issue-7