I also go by Chris, my English name. I'm a second-year PhD student in the Computing & Mathematical Sciences Division at MBZUAI, advised by Prof. Timothy Baldwin and Dr. Yova Kementchedjhieva.

My research asks a question that sounds simple but isn't: when a vision-language model answers, how much should we trust it? I work on uncertainty quantification and hallucination detection for multimodal models. I want to figure out whether a model failed to see something or failed to know it, and to catch the mistake before it reaches a user.

I'm a builder at heart, though. I like research that ends in something you can actually run, and I enjoy writing tools other people end up using.

Before MBZUAI, I finished my Computer Science degree at HCMUT (VNU-HCM) in three and a half years, working with Assoc. Prof. Tho T. Quan on machine learning, multimodal analysis, and Vietnamese and minority-language NLP. I still have a soft spot for that last one.

khang:
  now:       2nd-year PhD @ MBZUAI · Abu Dhabi, UAE
  advisors:  [Timothy Baldwin, Yova Kementchedjhieva]
  focus:     [vision-language models, uncertainty quantification, hallucination detection]
  previous:  [Visiting Scholar @ NUS (2026), B.Sc. Computer Science @ HCMUT]
  born:      2003

• 2026-08 Teaching Assistant for Mathematical Foundations for AI at MBZUAI: ran the lab sessions and wrote problems for the final exam.
• 2026-06 Visiting Scholar at NUS with Prof. Anji Liu, working on uncertainty and hallucination in VLMs.
• 2026 Program Committee member and reviewer for CSoNet 2026.

_{Papers live on Google Scholar. This page is for the code.}

v7
An AI-powered Vietnamese input method: type x0ch2 and get xin chào. You type only initial consonants and tones, and v7gpt, a small GPT with a custom tokenizer trained on Vietnamese text, uses context to fill in the rest. Runs on macOS and Windows. Built together with ducngg.
_{Python · PyTorch · nanoGPT · input method}

Synthetic Generator
Generates synthetic datasets that plug straight into ML pipelines, for when the real data is scarce, private, or just not there yet.
_{Python · ML pipelines · data generation}

sidecomments
A LaTeX package that puts review comments in margin boxes, with arrows pointing to the exact text, figure, table, or equation. Works with ACL/ARR, ICLR, CVPR, and AAAI templates.
_{LaTeX · TeX package · Overleaf}

VitalFit Persona
A multi-agent healthcare assistant. Specialized agents call tools and work together on healthcare questions.
_{Python · agentic AI · Redis · Google APIs}

MERIT
Causal episodic memory for LLM agents, so they learn from feedback and repair their own failures instead of repeating them.
_{Python · LLM agents · memory · SQL}

HCMUT Chatbot
An LLM and RAG chatbot that answers university admission questions.
_{Python · RAG · LLMs}

TI-JEPA
A text-image joint-embedding predictive architecture for multimodal fusion. It started as my capstone project, which was graded 9.9/10.
_{PyTorch · Transformers · multimodal}

HuTieuBERT
A morpheme-aware Transformer that adds explicit morphological structure to pretrained encoders for isolating languages like Vietnamese.
_{PyTorch · BERT · tokenization}

_{See all repositories →}