Radchaneeporn Changpun

I'm a graduate student in computer science at the Department of Computer Engineering, Chulalongkorn University in Bangkok, Thailand, under the supervision of Professor Peerapon Vateekul, Ph.D. with Titipat Achakulvisut, Ph.D. and Professor Arunya Tuicomepee, Ph.D. as co-advisors.

Currently, I am researching about Large Language Models focusing in the area of Low-resouce language LLMs, Finetuning, and Agentic Workflow

CV / Github / Linkedin / GoogleScholar / Email / Huggingface

Research

I'm interested in data science, machine learning, and deep learning focused on Natural Language Processing(NLP). Area of my research is about Large Language Models.

Publications

Agentic Stage-Based LLM Framework for Multi-Turn Mental Health Support Conversations in Thai
Radchaneeporn Changpun, Naphat Khoprasertthaworn, Pipat Jongpipatchai, Theerin Petcharat, Krittapas Rungsimontuchat, el al.
The 20 th International Joint Symposium on Artificial Intelligence and Natural Language Processing

Presented on November 2025, 14

Read my paper

We developed an agentic stage-based LLM framework that guides conversations through five counseling stages, drawing from Person-Centered Therapy and Acceptance and Commitment Therapy. Our system uses three specialized types of agents: stage based agent for each framework stage, another one approach selection agent selects appropriate counseling approaches, and monitoring agent manages stage transitions. The framework achieved a 79% positive user reaction rate, significantly outperforming baselines. Real user testing and evaluation by counseling practitioners confirmed improvements across seven of eight mental health support metrics, demonstrating potential for scalable LLM-based mental health support in Thailand.

Applied Artificial Neural Network for Hydrogen Sulfide Solubility in Natural Gas Purification
Prathana Nimmanterdwong, Radchaneeporn Changpun, Patipon Janthboon, el al.
Link to Paper

This publication is about applying a function in the MATLAB program and knowledge of Data Science to develop an artificial neural network model for predicting hydrogen sulfide solubility in natural gas purification processes. The model obtained a coefficient of determination (R2) of 0.9817 and a mean square error (MSE) of 0.0014.

Projects

	DMIND Chatbot LLM based mental health support chatbot in Thai. Currently, Dmind Chatbot is continue researching and developing under Center of Excellence in Digital and AI for Mental Health (AIMET) Radchaneeporn Changpun, Naphat Khoprasertthaworn, Pipat Jongpipatchai, Theerin Petcharat, Krittapas Rungsimontuchat Design initial architecture for chatbot framework Research and design multi-turn chatbot workflow LLMs with mental health support knowledge Coordinate with domain experts to design and improve the chatbot to meet domain practices Conduct experiments to find optimal techniques for the chatbot Led research and developer teams for the first phase of the project Led interns in designing and deploying the chatbot for the first phase of the project
	Multi-Label text classification Project Github I classify Scopus publications using encoder representation from transformers language model (RoBERTa), achieving a significant improvement of 40.3% in the Macro F1 Score (0.6687) compared to the baseline model (0.1894), demonstrating the effectiveness of transfer learning in enhancing text classification performance Implemented data preprocessing techniques, including tokenization, encoding, and data splitting, to prepare the dataset for training and evaluation Designed and developed a custom RoBERTa-based neural network architecture for multi-label classification, incorporating dropout regularization and a linear classification layer Utilized PyTorch and Hugging Face libraries to efficiently train and evaluate the model, leveraging GPU acceleration for improved performance
	ILabor- LLM based Thai Personal Income Tax Chatbot Project Github This group project conducted an empirical study comparing LLM techniques (agentic RAG, Naive RAG, Long Context LLM, Vanilla inference) for answering Thai Personal Income Tax (PIT) questions, measuring performance with automatic NLP metrics (BERT score, BLEU, ROUGE-L), LLM-as-a-judge, and qualitative analysis Preprocess Data (ความรู้ภาษีเงินได้บุคคลธรรมดา จากกรมสรรพากร: https://www.rd.go.th/62337.html) using OCR and Web scraping
	RAG Project Github I developed a RAG technique to improve the hallucination of Llama2-13B using the vector database created from Scopus publications Preprocessed the dataset for semantic indexing and utilized embedding model to generate semantic embeddings for the dataset Integrated Pinecone, a vector database, to store and efficiently retrieve relevant context Employed Meta's Llama-2-13b-chat-hf as the backbone for generating response and leveraged the Llama 2 tokenizer to preprocess and tokenize user queries and context for input to the LLM Utilized the Langchain framework to streamline the integration of the LLM, vector database, and embedding model Developed a RAG pipeline by combining the LLM with the retrieved context to generate accurate and contextually relevant responses to user queries, enabling the LLM to generate coherent and meaningful responses based on the provided context

My Blog

CS Graduate Degree Reflection
My blog about my experience and lessons learned during my computer science graduate degree at Chulalongkorn University, Bangkok, Thailand.
Read My Blog

Last update: January 17, 2026

This website is adapted from the source code from jonbarron's website.