Dr. Bao Gia Doan

I am an AI researcher and Postdoctoral Research Fellow at UNSW Sydney. I build and evaluate reliable language-model systems for specialist domains, particularly where labelled data and established benchmarks are scarce. My work spans retrieval-augmented generation, domain adaptation, synthetic data generation, neural information retrieval, and the evaluation of coverage, faithfulness, and hallucination.

My latest research introduces DoRA, a benchmark construction and evaluation framework for specialist domains, accepted as a main paper at EMNLP 2026. DoRA turns small collections of specialist documents into auditable training and evaluation data for RAG systems. Using defence as a high-stakes case study, it investigates how open-source LLMs can be adapted and evaluated under domain shift while remaining grounded in evidence.

My previous research focused on adversarial machine learning and the robustness of deep neural networks, including adversarial examples, data poisoning, backdoor attacks, and privacy-preserving techniques such as differential privacy and federated learning.

More broadly, I am interested in trustworthy AI deployment, efficient adaptation of open-source models, and robust evaluation for knowledge-intensive applications.

Current position

At UNSW Sydney, I develop reliable retrieval-augmented generation systems that connect large language models with external knowledge bases. My work covers semantic retrieval, prompt engineering, vector embeddings, neural information retrieval, fine-tuning, evaluation, and the scalability of knowledge-intensive applications.

Previous positions

Before joining UNSW, I completed my PhD and worked as a Postdoctoral Research Fellow at the University of Adelaide, focusing on the robustness of deep neural networks. Before academia, I spent nearly four years at Intel Vietnam as a Senior Process and Equipment Engineer.

For a detailed overview, see my CV.