I build grounded AI systems that connect models, data, tools, evaluation, and usable software.
LinkedIn · Live AI Demos · Email
I am a Computer Science graduate from Swinburne University of Technology Sarawak, with a double major in Artificial Intelligence and Cybersecurity.
I build complete AI applications that combine LLMs, retrieval, computer vision, backend services, databases, evaluation, testing, and user-facing interfaces.
I am open to graduate opportunities in AI Engineering, LLM Engineering, Machine Learning, and Software Engineering.
An evaluation-driven RAG system for asking technical questions about the FastAPI codebase.
The system searches source code, documentation, and resolved GitHub issues before generating an answer. It returns file, symbol, and line-level references so each response can be checked against the original source.
Key engineering work
- Built AST-aware code chunking with Tree-sitter
- Compared dense, hybrid, reranked, and query-rewritten retrieval
- Added grounded citations and evidence-aware refusal
- Created a labelled evaluation benchmark and regression pipeline
Results: Recall@5 91% · MRR 0.71 · Faithfulness 89% · Correctness 91% · Refusal accuracy 7/7
Stack: Python · LLMs · RAG · Streamlit · ChromaDB · BGE Embeddings · Tree-sitter · Pytest
An LLM-powered data analyst that investigates business questions against a PostgreSQL database.
The system plans an investigation, discovers the relevant schema, generates SQL, checks it for safety, executes it in a read-only transaction, validates the result, and explains the final findings.
Key engineering work
- Designed an eight-stage investigation workflow
- Added SQL safety inspection with SQLGlot and read-only execution
- Implemented automatic SQL repair and result validation
- Added token, latency, cost, and execution tracking
Project snapshot: 103 tests · ~650,000 database rows · 6 planted anomalies · 3 repair attempts maximum
Stack: Python · LLM Tool Calling · FastAPI · PostgreSQL · SQLAlchemy · SQLGlot · Docker · Pydantic
A multimodal AI system that combines concrete-defect detection with evidence-grounded repair guidance.
The vision model detects cracks, corrosion, and spalling. A LangGraph workflow then retrieves relevant information from USACE engineering guidance and uses that evidence to produce a cited response.
Key engineering work
- Fine-tuned YOLOv8 for three concrete-defect classes
- Built a LangGraph workflow with retrieval grading and query rewriting
- Added section-level citations and unsupported-question refusal
- Evaluated retrieval, citation validity, faithfulness, and refusal behaviour
Results: 1,770 images · 5,897 objects · Retrieval 28/28 · Citation validity 33/33 · Faithfulness 31/32
Stack: Python · YOLOv8 · LLMs · LangGraph · LangChain · ChromaDB · Gradio
A JADE-based routing system using master and delivery agents, genetic algorithms, simulated annealing, nearest-neighbour search, and local optimisation.
Stack: Java · JADE · Multi-Agent Systems · Genetic Algorithm · Simulated Annealing
A mobile-first campus platform for navigation, safety, student support, events, and administration.
I worked mainly on the admin console, emergency and exit-management workflows, support features, full-stack integration, and team coordination.
Stack: Next.js · React · TypeScript · Supabase · Tailwind CSS
AI & LLMs: Large Language Models, RAG, LangGraph, tool calling, embeddings, vector search, prompt engineering, computer vision, object detection, model evaluation
Backend & Data: Python, FastAPI, PostgreSQL, SQLAlchemy, ChromaDB, Pydantic, REST APIs, Docker
Software Engineering: Java, TypeScript, Next.js, React, Git, Linux, testing, CI/CD
Building AI systems that move from raw data to evidence, reasoning, and useful action.