Thodsaporn Chay-intr, D. Eng.

Head of AI at iApp Technology Co. Ltd., AI/ML Engineer, and NLP Researcher

Resume

Click here to download the PDF version of my resume.

Work Experience

  • AI/ML Engineer and Head of AI
    iApp Technology Co., Ltd., Thailand, Jan 2024 - Aug 2024
    • Developed an LLM-based chatbot/search using LLMs with llama.cpp, TensorRT-LLM, and RAG, achieving 97.67% QA accuracy and 87.53% perfect recall with response times under 6.5 seconds.
    • Synthesized training samples for text classification with LLMs, boosting accuracy above 90% and improving by up to 27%.
    • Managed and led a multidisciplinary team to develop AI projects from inception to deployment, including OCR, ASR, TTS, LLMs, and image processing, doubling team capacity and efficiency.
    • Led R&D initiatives to transform cutting-edge research into practical AI products.
  • ML Researcher and Lecturer
    Artificial Intelligence Association of Thailand, Thailand, Jan 2018 - Aug 2023
    • Delivered Python courses on data analysis and ML to hundreds of AI enthusiasts, covering tools such as pandas and scikit-learn, and algorithms including SVM and NNs.
    • Consulted with industry professionals and scholars on ML concepts, guiding plans and assisting in publishing over 20 international research papers and projects.
  • Research Assistant
    Tokyo Institute of Technology, Japan, Sep 2019 - Mar 2020
    • Collaborated with a multidisciplinary team to develop modules for NLU/NLG units for Japanese conversational dialogs.
    • Implemented a Seq2Seq model using BiLSTM with cross-attention in PyTorch for more natural text generation, validated through human evaluations.
  • ML Engineer and Researcher
    iApp Technology Co., Ltd., Thailand, Mar 2017 - Feb 2018
    • Constructed the first Thai Treebank with over 5,000 entries with a team of linguists and developers, enhancing resources, research, and applications in the Thai NLP community.
    • Developed a syntactic annotation tool in native (Java) and web (React, Python) applications, deployed on GCP, supporting ongoing resource developments.

Selected Projects

  • LLM-based Conversational AI System for General Banking Queries (Mar 2024)
    • Contributed to POCs and development of an LLM-based chatbot/search system for a tech innovation division of a leading bank in Thailand using LLaMa (2, 3, and 3.1), OpenThaiGPT, llama.cpp, TensorRT-LLM, and RAG.
    • Achieved 97.67% QA accuracy, 87.53% perfect recall, and maintained response times under 6.5 seconds.
    • Introduced intent classification with over 92% accuracy as a guardrail to filter the input/output of the system, ensuring the responses align with banking policies.
  • Extreme Fine-tuning: A Novel and Fast Fine-tuning Approach for Text Classification (EACL 2024) (Mar 2024)
    • Proposed a novel text classification fine-tuning approach incorporating backpropagation with extreme learning machine, reducing fine-tuning time while retaining classification accuracy and F1-score.
    • Attained faster fine-tuning time by up to 74.8% with comparable scores over recent state-of-the-art models on MELD, IEMOCAP, IMDb, and AG News datasets.
  • LLaVAC: Fine-tuning LLaVA as a Multimodal Sentiment Classifier (Jan 2024)
    • Proposed a method to fine-tune Large Language-and-Vision Assistant (LLaVA) as a classifier for classifying multimodal sentiment labels by designing a prompt to consider unimodal and multimodal labels and generating predicted labels.
    • Outperformed state-of-the-art baselines by up to 7.31% in accuracy and by 8.76% in weighted-F1 in the MVSA-Single dataset.
  • A Unification-based Knowledge Graph Construction for Thai Profile Generation from Online Resources (Sep 2023)
    • Constructed a knowledge graph for Thai researchers, using 6+ million entries crawled from online research databases.
    • Designed a semi-supervised method with multi-task learning to extract entities/relations, improving F1-score by 8% over baseline.
  • Simple2In1:A Simple Method for Fusing Two Sequences from Different Captioning Systemsinto One Sequence (Sep 2023)
    • Developed a T5-based generative model for Thai captions fusion, outperforming baselines by 5.2% in sBLEU and ROUGE-L scores.
    • Accomplished a sBLEU score of 79% and a ROUGE-L score of 90% for a small captioning dataset comprising 3,168 samples.
  • LATTE: Lattice ATTentive Encoding for Character-based Word Segmentation (Jun 2023)
    • Proposed a sequence labelling method that integrates multi-granularity linguistic units, Lattices, GNNs, PTMs, and Attention Mechanism to generate and refine text representations for word segmentation. in PyTorch with PyG.
    • Achieved state-of-the-art performance (97.7% to 99.4% of F1-score) across Asian languages: Japanese, Chinese, and Thai.
  • Multimodal Sentiment Analysis Using Multiple Labels from Different Modalities (Mar 2023)
    • Collaborated with students to design and implement a sentiment analysis model for social network data, leveraging text, image, and multimodal labels using CLIP, BERT, and RoBERTa. Yielded up to 2% improvement in F1-score over recent models.
    • Attained F1-scores of 74.1% for MVSA-single and 62.0% MVSA-multiple datasets.
  • Detecting Fraud Job Recruitment Using Features Reflecting from Real-world Knowledge of Fraud (Mar 2022)
    • Developed a method to classify fake job recruitments using a set of novel features designed to reflect fraudster behaviors.
    • Yielded accuracy of 97.64% for Employment Scam Aegean Dataset (EMSCAD).
  • Public Budget Usage Monitoring System (Feb 2019)
    • Cooperated with an interdisciplinary team to develop a monitoring system that utilizes Scrapy to crawl large-scale unstructured data from government sites, such as procurement and budget portals, for corruption detection in text data. Deployed by two organizations.
    • Developed a text classification method in TensorFlow, with rule-based enhancements, for corruption detection, validated by experts.
    • Obtained Bronze Medal in The 47th International Exhibition of Inventions Geneva

Education

  • Doctor of Engineering, Tokyo Institute of Technology, Japan, 2019 - 2023
  • Master of Engineering, Sirindhorn International Institute of Technology, Thammasat University, Thailand, 2015 - 2018
    • Major: Information and Communication Technology for Embedded Systems
    • Advisor: Prof. Thanaruk Threeramunkong
    • Thesis: A Framework of Thai Treebank Construction and Grammar Refinement
    • Scholarship: TAIST - Tokyo Tech Scholarship
  • Bachelor of Science, Thammasat University, Thailand, 2011 - 2015
    • Major: Computer Science
    • Advisor: Dr. Ratchata Peachavanish
    • Thesis: Web Application for Learning Thai Language using Song Lyrics
    • Activity: President of Computer Science, Faculty of Science, Thammasat University

Skills

  • Technical Skills
    • Programming Languages: Python, C/C++, Rust, Java
    • ML Toolkits: PyTorch/Lightning, TensorFlow, Hugging Face, PyG, OpenCV, Scikit-learn, Spacy, NLTK, llama.cpp, TensorRT-LLM
    • Tools & Technology: Linux, Hadoop/Spark, SQL, NoSQL(MongoDB, Neo4j), Docker, Elasticsearch GCP, AWS, Git
  • Languages
    • Thai: Native
    • English: Professional working proficiency
    • Japanese: Limited working proficiency