- 🔭 I build end-to-end ML systems: from data collection & annotation → fine-tuning → evaluation → serving in production.
- 🧾 Current focus: Key Information Extraction (KIE) from documents with vision-language models (Florence-2, LoRA, continual learning).
- 🇻🇳 Interested in Vietnamese document understanding and low-resource datasets.
- 🌱 Currently learning:
- 📍 Based in Vietnam · Open to internships / full-time roles in AI/ML.
Multi-domain receipt information extraction (company, date, address, total) across SROIE (EN), MC-OCR (VI) and a self-collected Vietnamese supermarket dataset.
| 🎯 Accuracy | 88.21% macro ANLS (+7.7 pts over no-replay baseline) |
| 🧠 Continual learning | Weighted replay reduces forgetting — macro BWT −1.13 |
| ⚡ Efficiency | LoRA trains only 1.87% of params (5.17M / 277M) |
| 🏎️ Serving | ~585 ms/image on T4, ~1.4 GB VRAM · FastAPI + Docker |
| 🛡️ Robustness | Tolerant JSON parser → ~99–100% valid structured output |
PyTorch PEFT/LoRA Florence-2 FastAPI Docker Streamlit Hugging Face
➡️ Live demo on Hugging Face Spaces · Source code
Languages
ML / Deep Learning
Serving & MLOps
“Ship models that work outside the notebook.”



