Projects

ETOM: A Five-Level Benchmark for Evaluating Tool Orchestration within the MCP Ecosystem EACL Findings ’26

Constructed a five-level benchmark with a 2,000+ tool pool and Equal Function Sets to objectively evaluate multi-turn LLM agent tool orchestration in hierarchical MCP ecosystems.

AIResearchLLMs

Academic Project Page: An Agent Skill for Research Websites

Built an agent skill that turns papers, figures, and videos into responsive academic project pages, with source checks and browser-based audits for content, accessibility, and layout.

AIResearchLLMs

LeetCode Companion: A Local-First Study Dashboard

Built a local-first study dashboard and Chrome extension to plan LeetCode practice, schedule reviews, and track solving sessions and submissions.

Web DevelopmentChrome Extension

Call2Care: Voice-First AI for Elder Care

Built a voice-first elder-care prototype combining LLMs, speech recognition, and RAG over health-education documents. Streamed check-in transcripts, AI-generated risk flags, and call summaries to a live caregiver dashboard using FastAPI and WebSockets.

AIRAGVoice AI

Multi-Objective RL for Balancing Power and Performance in 5G

Developed a preference-conditioned RL controller for 5G bandwidth switching to balance power, delay, and packet loss. Achieved 0.236 vs. 0.042 Pareto hypervolume against a heuristic baseline on unseen stored-streaming traffic in simulation.

AIReinforcement Learning

Developing an AI System for Smart Traffic: From Vehicle Detection to Data Analytics Dashboard CPEG FYP Champion

Fine-tuned YOLO11 and ByteTrack for real-time multi-vehicle detection and tracking, integrated with a FastAPI and Streamlit dashboard for live traffic analytics.

AIComputer Vision

Serverless vs. Server-Based Image Classification Deployment

Benchmarked serverless (OpenFaaS on Kubernetes) versus server-based (Flask) architectures deploying MobileNetV2, evaluating inference latency, throughput, and resource trade-offs under varying loads.

Cloud ComputingAI

Predicting Horse Racing Outcomes Using Machine Learning

Trained multi-layer perceptron and random forest models combined with a pointwise Learning to Rank (LTR) framework to predict race placement probabilities.

AIMachine Learning

TV Sensor

Engineered an ESP32-based ambient light sensing system using a VEML6040 sensor to detect TV power states with 96.77% accuracy in real time.

Embedded SystemsIoT

AdaIN Style Transfer and Image Classification

Implemented an Adaptive Instance Normalization (AdaIN) style transfer model and trained a VGG19 network with style-augmented data for robust cross-style image classification.

AIComputer Vision

Autonomous Line-Following Car

Built an autonomous Arduino robotic vehicle using infrared sensor arrays and custom decision-making logic for real-time track navigation.

Embedded SystemsRobotics

Data Exploration to Performance Optimization for Machine Learning

Built and benchmarked end-to-end ML regression and classification pipelines using feature engineering, neural networks, and grid search hyperparameter tuning.

AIMachine Learning