Technical University of Darmstadt

BMBF-funded project InsightsNet, acting as the team's NLP engineer to build production ML systems for knowledge extraction from scientific publications.

• Architected, developed, and deployed an end-to-end PDF-to-structured-data pipeline that processed over 100K scientific publications, incorporating layout analysis and entity-relation extraction.
• Developed novel layout-aware text editing model for PDF-to-Markdown conversion achieving 1.7× inference speedup over baseline models.
• Built citation intent classification system using orthogonal representation learning, improving F1-score by 8% over SOTA.
• Designed and implemented annotation guidelines, knowledge graph schema and quality control framework for a multi-disciplinary corpus, leading a team of 15 annotators and constructing an enterprise-ready knowledge graph with over 50K verified entity relations.
• Collaborated closely with domain experts, data owners, and software engineers to align knowledge graph modeling decisions with downstream analytics and application requirements.

Robert Bosch GmbH

Developed deep learning models for time-series forecasting in IoT sensor data (e-Bike sensors).

• Designed and benchmarked multiple generative models (GRU-VAE, conditional GAN) for multivariate sensor prediction, achieving ~45% MAE improvement and ~20% faster inference over baseline LSTM.
• Collaborated with hardware engineering team to integrate models into production prototype system.
• Delivered technical documentation and model handoff enabling seamless production deployment by engineering team.

Technical University of Darmstadt

Dissertation: Enhancing Scholarly Document Accessibility and Analysis

University of Stuttgart

Thesis: Semi-supervised Event-centered Emotion Analysis and Performance Prediction (in collaboration with Robert Bosch GmbH)

Henan Normal University

Layout-Aware Text Editing for Efficient Transformation of Academic PDFs to Markdown

Changxu Duan, ICDAR 2025 (top-tier Document Analysis conference) [GitHub] [Paper] [arXiv]

Semantically Orthogonal Framework for Citation Classification: Disentangling Intent and Content

Changxu Duan, Zhiyin Tan, TPDL 2025 (flagship Digital Libraries conference) [GitHub] [Paper]

Multi-Disciplinary Dataset Discovery from Citation-Verified Literature Contexts

Zhiyin Tan, Changxu Duan, JCDL 2025 (top-tier Digital Libraries conference) [Web Page] [Paper] [arXiv] [Code]

KG-Augmented LLM Hallucination Reduction

Poster, DTU MLOps Summer School 2022. Built entity-linking pipeline integrating Knowledge Graphs with LLMs to verify factuality and reduce hallucinations.

SPARQL Auto-Completion VS Code Extension

Open-source developer tool with 1K+ installs providing SPARQL IntelliSense, prefix completion and query authoring for knowledge graphs. [GitHub]