{"id":131,"date":"2025-09-05T00:30:44","date_gmt":"2025-09-04T23:30:44","guid":{"rendered":"https:\/\/soc2.ucd.ie\/?p=131"},"modified":"2025-11-25T14:06:24","modified_gmt":"2025-11-25T14:06:24","slug":"smartintent","status":"publish","type":"post","link":"https:\/\/soc2.ucd.ie\/index.php\/2025\/09\/05\/smartintent\/","title":{"rendered":"SmartIntent"},"content":{"rendered":"\n
\"\"
Overall architecture of SmartIntent<\/figcaption><\/figure>\n\n\n\n

Current smart building systems often face challenges such as complex device linkage and limited natural language understanding. Our research lab addresses these with\u00a0SmartIntent<\/strong>, an innovative, intent-recognition-based smart building control system. It aims to build a system that allows users to perform multi-language, multi-device, and multi-mode control and automation settings through natural language.<\/p>\n\n\n\n

The Power of Serverless and LLMs<\/h2>\n\n\n\n

SmartIntent\u2019s robust foundation lies in its microservice architecture and Knative Serverless deployment, combined with Large Language Models (LLMs) to achieve semantic parsing, context understanding, and device control command generation.<\/p>\n\n\n\n

Serverless Scalability<\/h3>\n\n\n\n

The system adopts a microservice architecture and Knative Serverless deployment, which greatly improves scalability and elasticity. This ensures zero operation and maintenance delivery, high availability, and on-demand resource scheduling capabilities.<\/p>\n\n\n\n

LLM-Powered Intent Understanding<\/h3>\n\n\n\n

The LLM Intent Parsing Module is central to SmartIntent\u2019s intelligence. It converts natural language commands into structured JSON instructions or automation rules. We\u2019ve utilized advanced LLMs like Grok 3 Beta for their strong performance in knowledge-intensive reasoning, long-context comprehension, and broad multi-task generalization, crucial for interpreting diverse user instructions.<\/p>\n\n\n\n

Our research found that providing a small number of input\/output examples (Few-shots) significantly improved model accuracy, while adding more (Many-shots) led to diminishing returns and increased token overhead. Fine-tuning with a high-quality, compact dataset of around 200 Chinese instructions achieved near-maximal accuracy and showed strong cross-lingual generalization, significantly reducing data annotation and training costs for multilingual tasks.<\/p>\n\n\n\n

SmartIntent in Action: Orchestrating Your Space<\/h2>\n\n\n\n

The middleware layer, described as the system\u2019s \u201ccore processing unit,\u201d coordinates user intent with device operations. It comprises:<\/p>\n\n\n\n