At Vidi Labs Ltd., I lead development of an indoor navigation system for visually impaired travellers at Hong Kong International Airport. It combines LiDAR mapping, visual localization and spoken directions; in trials with a blind user, it guided them to the destination gate in 98% of runs.
I designed the visual localization pipeline using NetVLAD retrieval, LightGlue matching and multi-frame PnP/RANSAC, reaching a median error of 1.3 cm / 0.22° on held-out frames. I also reduced post-pan latency from over 20 seconds to about 3 seconds by streaming frames for incremental processing and keeping models and map features resident on the GPU.
At Vidi Labs, I developed a landmark-based guidance system for HKIA that served 20–25 concurrent users on about 4 GB of VRAM. I also built a real-time translation backend for SOLOS smart glasses and a locally deployed chatbot optimized for on-device inference.
Previously, at Genuity Systems Ltd., I developed conversational AI and speech systems, including a RASA customer-service chatbot with LLaMA3 fallback handling and a WebRTC transcription integration. At Animo.AI, I developed real-time object detection and tracking services and a custom DenseNet-based pattern recognition model.
