Portfolio · LinkedIn · Hugging Face · Email
Final-year AI Engineering student at Xiamen University Malaysia. I build computer vision, generative AI and efficient deep learning systems end to end: model, API, UI, tests and CI.
Focus: efficient object detection · visual anomaly detection · RAG, LLM fine-tuning and evaluation · AI agents (MCP) · reinforcement learning
Open to: AI/ML internships (3–5 months from 29 March 2027) and project collaboration
| Project | What it is | Result | Links |
|---|---|---|---|
| SumProof | Receipt reading with a fine-tuned vision-language model, plus cited answers from 12,013 pages of real SEC filings | Receipt field F1 0.79 → 0.94 · Recall@5 0.53 → 0.71 · correct answers 8% → 35% · calculations 38% → 68% with budget-forced reasoning | GitHub · Hugging Face · Demo |
| DefectRAG | Defect detection from normal images only, plus a RAG assistant that cites its sources | 0.971 AUROC on VisA (from 0.950) · 0.992 on MVTec AD · answers with correct sources 42% → 89% · 500+ tests | GitHub · Hugging Face · Demo |
| PaddyLiteX (thesis) | Lightweight YOLOv8n paddy disease detector, 9 classes | 81.67% mAP@0.5 · −68% parameters · −73% GFLOPs | GitHub · Hugging Face · Demo |
| MarineLite | Underwater organism detection (DU-MobileYOLO + GhostConv + SimSPPF) | mAP@0.5 74.23% → 75.36% · parameters 4.70M → 4.37M | GitHub · Hugging Face · Demo |
| dPickleball | PPO agent in Unity ML-Agents (team project) | Team 2nd of 13 · my agent in the team's top 4 | GitHub · Hugging Face · Video |
More projects (coursework): Res-UCN, a lightweight ResNet-based CNN (GitHub · Hugging Face · Demo) · FER2013 facial expression recognition (GitHub · Hugging Face · Demo) · Alzheimer's MRI classification (GitHub · Hugging Face · Demo)
Render demos can take about a minute to wake. DefectRAG VisA figures are descriptive, not independent validation.
| Languages | |
| ML & Computer vision | |
| GenAI & NLP | |
| Data | |
| Web & APIs | |
| Cloud & Deployment | |
| Tools & Automation |