Mechatronics Engineering student at Universiti Teknologi Malaysia, working on applied AI for high-stakes systems — where a wrong answer has a cost, and the model has to explain itself.
Research interests: LLM reasoning and test-time compute, agentic systems, industrial fault diagnosis, and clinical decision support.
Test-Time Compute and Semantic Safety Filtering for Industrial Fault Attribution Accepted — IEEE IRAI 2026
An LLM diagnostic pipeline for industrial machinery that reasons over multiple fault hypotheses before answering, then passes every recommendation through a rule-based safety shield. Evaluated on the CWRU bearing dataset. Paper · Code
- Contributor to Mantis_SDK — MIT CompBio Group (Kellis Lab), via the MIT AI Mantis Rising Scholars program
- Robotics integration on the AIROST × DF Cobot project — ROS 2, MoveIt2, Gazebo
- Former Head of AI/ML, GDG on Campus UTM (Nov 2025 – Feb 2026)
Asclepios AI — MIT Emerging Talent Program Core AI engines for a clinical decision-support system in substance use disorder treatment, built on 1.4M patient records. Twin-engine segmented XGBoost framework plus a Chronicity Proxy algorithm that ranked among the top three predictors of relapse risk (AUC 0.75, TEDS).
ExoDetectAI — NASA International Space Apps Challenge End-to-end pipeline for automated exoplanet validation across Kepler, K2 and TESS. Ensemble of XGBoost, LightGBM and MLP with LEO-vetter and lightkurve, deployed as a Streamlit prototype.
Neuro-Symbolic Fake News Detector Hybrid LSTM and symbolic-logic model, improving interpretability and adversarial robustness over pure deep-learning baselines.
| Domain | Tools |
|---|---|
| Languages | Python, C/C++, ARM Assembly, SQL |
| ML / DL | PyTorch, Hugging Face, scikit-learn, XGBoost, LightGBM |
| LLM / Agents | sentence-transformers, chain-of-thought reasoning, tool-augmented agents, semantic retrieval |
| Robotics & CV | ROS 2 (Jazzy), Gazebo Harmonic, MoveIt2, ros2_control, OpenCV, Open3D |
| Tooling | Git, Linux, Docker, Streamlit |