Projects

Traffic Detection System

CARLA, Vision-Language Models, Edge AI GitHub | Demo

Developed a multimodal traffic detection system combining RGB and LiDAR sensor data to train a Vision-Language Model (VLM) for corner-case hazard detection in autonomous driving. Generated synthetic datasets using CARLA, applied RGB-LiDAR projection fusion, and optimized the trained model for real-time embedded inference using Raspberry Pi.

Designed and implemented a multimodal traffic detection pipeline that integrates RGB camera and LiDAR point cloud data for robust corner-case hazard understanding in autonomous driving scenarios. Built a synthetic dataset generation engine in CARLA, covering rare events such as occlusions, adverse weather, near-miss collisions, and unusual agent behaviors.

Developed an RGB-LiDAR projection-based fusion module to convert LiDAR depth and reflectance into RGB-aligned channels, enabling downstream Vision-Language Models (MiniCPM-o, Qwen-VL, BLIP-2 variants) to perform spatial reasoning and semantic grounding.

Implemented ontology-driven annotation using AVCCO ontology and TTL triples to guide model reasoning and generate structured hazard explanations. Optimized the full pipeline for real-time inference on embedded platforms using Python, PyTorch, and C++, including quantization, dynamic resolution scaling, and an edge-cloud hierarchical architecture where a lightweight edge model filters scenes before offloading complex reasoning to a larger backend VLM.

Lean Improvement Knowledge Library Platform

Enterprise Web System, Manufacturing IT

.NET 8, SQL Server, Dapper, JavaScript

Developed an enterprise-scale Lean Improvement Knowledge Library to digitize, standardize, and govern continuous-improvement cases across a manufacturing organization, supporting ~900 internal users as a centralized knowledge and analytics platform.

Designed and implemented a full-stack web system using ASP.NET Core (.NET 8) and SQL Server, with secure cookie-based authentication, role-based access control, and RESTful APIs to support long-term operational use.

Modeled organizational ownership through departments and positions instead of individual employees, enabling scalable maintenance and reducing administrative overhead during personnel or structure changes.

Built high-performance search, filtering, and analytics workflows using Dapper-optimized SQL queries, delivering sub-second response times and reducing improvement retrieval from minutes to seconds.

Implemented dynamic frontend components for cascaded classification, attachments, and in-browser preview, improving data consistency, usability, and cross-department reuse of Lean improvement knowledge.

Light Sensor Swarm System

ESP8266, Raspberry Pi, IoT Networking GitHub

Designed a distributed IoT system with six ESP8266 light sensor nodes communicating via UDP to a central Raspberry Pi hub. Built a dynamic discovery protocol reducing setup time by 50% and implemented power-aware sleep scheduling to extend battery life. Achieved 95% transmission reliability with real-time light tracking.

Real-Time Canny Edge Detection Optimization

OpenCV, Multi-threading, PSNR Evaluation

Enhanced real-time edge detection performance by optimizing Canny Edge algorithms with multi-threading and dynamic frequency scaling. Integrated PSNR-based image quality evaluation and benchmarked performance across different threading strategies for embedded applications.

Network Simulation

Python, FRRouting, ContainerLab, Docker

Designed and automated deployment of 10+ virtual routers using ContainerLab with dynamic BGP and OSPF routing protocols, reducing topology setup time by 40%. Coordinated a team of three to implement, test, and present the simulation, earning a positive faculty evaluation. The project emphasized scalable network automation, containerized lab orchestration, and hands-on routing protocol configuration in a virtualized environment.

Waybar and Omarchy Linux Workflow Setup

Linux, Hyprland, System Automation

Customized a minimalist Linux desktop environment using Hyprland and Waybar for keyboard-driven productivity. Integrated VPN toggles, multi-clock modules, and fcitx5-Rime support, refining workflow aesthetics with MapleMono fonts and script-based automation.