Projects

Vehicle-Integrated Photovoltaics (VIPV) present unique forecasting challenges that differ fundamentally from fixed-site solar systems. This research develops a VIPV-specific data pipeline and benchmarks machine learning models — on two years of real-world BMW i3 sensor data collected in Sydney, Australia.

A key contribution is the causal day-shape prototype feature engineering framework, which improves CatBoost's RMSE gain from −0.48% to +15.14% across walk-forward validation. Interpretability analysis using SHAP and TFT's Variable Selection Network reveals that short-term VIPV forecasting is driven more by operational state than by raw irradiance — a finding with direct implications for onboard energy management system design.

Can AI faithfully represent divergent philosophical interpretations? This project develops a custom RAG system to explore LLM understanding of consciousness-related questions through a translation-conflict-aware evaluation framework.

Evaluated across 11 translations of an ancient philosophical classic, using BERTScore and n-gram analysis to measure how responses shift as new sources are incrementally added. Pipeline integrates query expansion, cross-encoder reranking, and consciousness-specific prompt engineering.

This project implements a multi-sensor autonomous robot integrating LiDAR and depth camera for real-time ball detection and autonomous kicking. The perception pipeline fuses point cloud and RGB-D data to achieve robust object localization across varying lighting conditions.

The project addresses core challenges in embodied AI: sensor noise handling, multi-modal data fusion, and bridging detection confidence with reliable physical action execution. A key research direction combines traditional SLAM-based navigation with a reinforcement learning policy, enabling the robot to improve kicking behavior through environmental interaction rather than hand-crafted rules.