I am a final-year Computer Science student transitioning into Software & Data Engineering. My engineering philosophy centers on deconstructing complex system problems into well-defined, atomic sub-problems, establishing rigorous data contracts, and grounding solutions in verified academic and scientific research rather than reinventing the wheel.
As an AI-native engineer, my core strengths lie in System Analysis, Architectural Design, and Data Pipeline Orchestration - leveraging state-of-the-art AI agents to accelerate high-velocity, robust code execution.
- Data Engineering & Knowledge Graphs: Ontology modeling, structured data pipeline execution, and semantic extraction.
- Document Intelligence & NLP: Information extraction, representation learning, and hybrid lexical-semantic search.
- Computer Vision & Automation: Multi-stage vision pipelines and robust data validation architectures.
| Category | Technologies & Tools |
|---|---|
| Languages | Python (Core), SQL, C++, C# |
| Data & MLOps | MongoDB, GraphDB, RDF/SPARQL, Prefect, DVC, Weights & Biases, Pandas |
| AI / NLP / CV | PyTorch, Ultralytics (YOLOv11), OpenCV, RapidOCR, Sentence-BERT, Hugging Face |
| Tools & Dev | Git, GitHub, Docker, FastAPI, Streamlit, Gradio, Pytest |
- Context & Architecture: Designed a modular 6-stage, 13-module data pipeline to automatically ingest, clean, and transform 14,039 NeurIPS papers (2021-2024) into a structured knowledge graph.
- Key Implementations: Integrated LLM-based semantic extraction (Qwen via vLLM) for citation tracking and author-affiliation mapping. Built a fault-tolerant execution engine with MongoDB-based isolated data blocks to recover seamlessly from API limits. Exported 2.1M RDF triples to GraphDB.
- Stack: Python, vLLM, MongoDB, GraphDB, Prefect, SPARQL
- Context & Architecture: Engineered a 2-tier detection pipeline designed to audit student theses by combining lexical matching and deep semantic analysis.
- Key Implementations: Implemented lexical matching (BM25 + TF-IDF char n-grams) paired with a semantic layer (Sentence-BERT). Optimized infrastructure costs by dynamically routing only unmatched sentences to the semantic layer, cutting processing time by 13%.
- Stack: Python, Sentence-Transformers, Rank-BM25, Scikit-learn, FastAPI
- Context & Architecture: Engineered a 5-stage CV pipeline with fail-fast logic and Pydantic type safety to extract ISO 6346 container IDs from uncontrolled real-world logistics environments.
- Key Implementations: Deployed YOLOv11 for object and keypoint localization. Structured a BRISQUE+SVM quality gate to reject degraded images early, paired with a hybrid OCR module (RapidOCR/Tesseract) and algorithmic check-digit validation.
- Stack: Python, PyTorch, YOLOv11, OpenCV, RapidOCR, DVC, Weights & Biases
- Email: duyhuu284@gmail.com
- LinkedIn: linkedin.com/in/duyhxm
- GitHub Personal: github.com/duyhxm