AI Engineer | AI Researcher | Data Scientist
M.S. candidate in Artificial Intelligence and Robotics at Sejong University
I build applied AI systems that connect research ideas with working products. My current research at Sejong University's Vision Language Intelligence Lab focuses on vision-language models, GUI grounding, and computer-use agents, especially training-free refinement methods for dense screen understanding.
Before graduate research, I worked as a data scientist and AI engineer on production-style systems: computer vision attendance, VLM-based CCTV filtering, time-series forecasting, web behavior analytics, and large-scale data enrichment pipelines.
- Vision-language models for GUI grounding and computer-use agents.
- Training-free adaptive refinement using attention maps and region selection.
- GraphRAG and knowledge-graph question answering over multimodal technical documents.
- Computer vision pipelines for detection, tracking, video understanding, and automation.
- Forecasting and customer analytics systems built from messy real-world data.
| Work | Role | Status / Result |
|---|---|---|
| DART: Density-Aware Adaptive Refinement | First author | IEEE Access under review; improved ScreenSpot-Pro accuracy from 73.0% to 80.9% on KV-Ground-8B |
| Training-Free ROI Selection via Semantic Vector Fields | Co-author | Manuscript revised for AAAI 2027 submission |
| CGAR: Confidence-Guided Adaptive Refinement | First author | Research project and patent submission for adaptive GUI grounding |
| A Dataset for Computer-Use Agents | First author | Dataset and benchmark project in preparation |
| Repository | What it shows |
|---|---|
| Density-Aware-Adaptive-Refinement-Technique | VLM-based GUI grounding, attention-map analysis, adaptive inference-time refinement |
| CCTV-Video-Filtering | VLM video classification with Qwen3-VL and InternVL3; 94.44% accuracy on a manually labeled test set |
| Structural-Drawing-GraphRAG | Hybrid GraphRAG for construction drawings; extracts structured facts with a VLM, links entities across sheets, imports them into Neo4j, and answers questions with source-sheet evidence |
| Vision-based-Communication | YOLOv11 + ByteTrack pipeline for vision-guided beamforming; +20-29% average data-rate improvement |
| Electricity-Consumption-Forecasting | STL-based monthly forecasting for 13 provinces; MAPE below 6% across all provinces |
| Restaurant-Traffic-Forecasting | XGBoost forecasting for 440+ restaurants; reduced WAPE from 25.3% to 22.0% |
| Face_Recognition | Real-time face-recognition attendance system with SSD, tracking, embeddings, and cosine matching |
| log-web-fshop | 96M+ web logs processed into customer journeys, behavioral segments, and exit-intent signals |
| Google-Maps-Business-Scraper | Large-scale business data crawling and enrichment from Google Maps, Chotot, and LinkedIn sources |
AI / ML: Vision-language models, GUI grounding, computer-use agents, GraphRAG, knowledge-graph QA, computer vision, object detection, object tracking, face recognition, deep learning, CNNs, classification, time-series forecasting.
Programming: Python, PyTorch, OpenCV, FastAPI, Pandas, NumPy, PySpark, C/C++, MATLAB.
Tools: Git, Linux, Jupyter Notebook, Neo4j, Docker.
Languages: Vietnamese native, English fluent, Korean basic.
- M.S. in Artificial Intelligence and Robotics, Sejong University, Seoul, Korea.
- Double Degree M.S. in Convergence Engineering for Intelligent Drone, Sejong University.
- B.E. in Mechatronics Engineering (PFIEV), VNU - Ho Chi Minh University of Technology, Vietnam.
I am open to AI engineer, data scientist, and research engineer opportunities involving vision-language models, computer vision, forecasting, and applied automation.
- Email: lexuanhoang120@sju.ac.kr
- LinkedIn: xuan-hoang1999
- Portfolio: lexuanhoang120.github.io
- Resume: Le_Xuan_Hoang_Resume.pdf