PROJECTS

An end-to-end framework for building operation and maintenance, integrating a BMS-linked data model, a site-adapted O&M knowledge base, and an LLM-based search agent.

The platform interprets runtime context and operator queries to automatically retrieve equipment data and execute diagnostic tools, generating traceable FDD workflows.

Deployed in a Hong Kong government building, the system successfully demonstrated the diagnosis of a chiller compressor alarm.

We investigates the challenge of adapting Time-Series Foundation Models (TSFMs), like IBM Granite and Amazon Chronos, to Building Energy Forecasting (BEF) tasks, where it is demonstrated that straightforward fine-tuning yields limited performance gains.

To address this, the authors propose a novel Contrastive Curriculum Learning (CCL) method that organizes training data by difficulty, utilizing contrastive learning to measure the complexity of simulated data relative to real-world samples.

An intelligent document parsing solution for the industrial energy sector, utilizing a Vision-Language Model (VLM) to interpret and structure technical manuals for HVAC and power equipment.

By leveraging reinforcement learning methods (such as GRPO, Reinforce++, RLOO, and DAPO) to optimize VLM performance in parsing industrial system operation documentation, and by designing reasoning-based prompt formats alongside corresponding reward calculation schemes, we achieved a general performance improvement of 7-8% compared to baseline fine-tuning approaches.

An on-device face detection and recognition system for Android tablets built on the Rockchip RK3399, delivering millisecond-level identity verification with all matching performed locally.

The pipeline pairs a lightweight MobileNetV3-based RetinaNet detector with a ResNet-34/50 recognition backbone trained under ArcFace metric learning, balancing speed against accuracy within a tight edge compute budget.

Inference was optimised on NCNN/MNN with NEON acceleration, operator fusion and FP16/INT8 quantisation, while infrared liveness detection guards against photo, video and mask spoofing. The system serves access control, attendance and self-service verification terminals.

An approach that transfers knowledge from pre-trained time-series foundation models into compact, physically interpretable RC models of indoor thermal dynamics, ready to drive model predictive control.

The solution was deployed on a VRF air-conditioning system in an occupied meeting room at Osaka University, integrating real-time building data, BACnet equipment control, a monitoring interface and a workflow-driven control backend.

Evaluation across hundreds of real buildings showed a 23.9% improvement in thermal dynamics prediction accuracy, and over 30 days of closed-loop operation across both cooling and heating modes delivered more than 9.2% energy savings.