Teerapol Saengsukhiran Wintoplay

Hi 👋, I'm Teerapol Saengsukhiran (aka Win)

💸 Ex-YSTP scholarship recipient (maintaining GPA > 3.00 at all cost)
🎓 Education: M.Eng. in Computer Engineering (Chulalongkorn University, 2025–present) | B.Eng. in Computer Engineering (SIIT, 2020–2024) – GPAX 3.73
🚀 Aspiration: ML Researcher / ML Engineer
🌎 English: C1 (IELTS 7.0)
📚 Recent achievement: First author at ACL 2026 (Poster) – Evaluating Perspectival Biases in Cross-Modal Retrieval
🧠 Experience: NLP (OpenThaiGPT team, Agile), Time series classification, tabular data modeling, object detection, MLLMs, bias evaluation, diffusion models

Chulalongkorn University – M.Eng. Computer Engineering (Aug 2025 – Present)
Sirindhorn International Institute of Technology (SIIT) – B.Eng. Computer Engineering (Aug 2020 – May 2024)

First author of Evaluating Perspectival Biases in Cross-Modal Retrieval, accepted to ACL 2026 (Poster)
Co-developed DLBKL (Discounted Language Bias KL-Divergence) to quantify prevalence bias in image-to-text retrieval
Co-designed triplet-based evaluation set to distinguish culturally associated vs. semantically faithful retrieval
Identified biases and failure modes in modern MLLM-based retrievers
Delegated tasks, managed concurrent work, aligned advisors, created feedback loops

Trained Multimodal Large Language Models (MLLMs) for Thai cultural context understanding
Enhanced Thai OCR capabilities for complex document layouts
Co-designed synthetic data pipeline and optimized for local-language nuances
Customized training pipelines and hyperparameters for early-stage MLLM stability

Researched efficient fine-tuning methods (PEFT) for LLMs
Proposed QLoRA as standard practice for fine-tuning causal LLMs under compute constraints
Trade-off exploration between QLoRA and full fine-tuning (performance, memory, time)
Learned SLURM system of Lanta supercomputer

Curated datasets for sport object detection and trained YOLO models
Designed web interface for inference using Gradio
Implemented signal classification models using 1D-Convolution and Transformer architectures

Evaluating Perspectival Biases in Cross-Modal Retrieval – ACL 2026 (Poster) [arXiv]
Investigated prevalence bias (linguistic) and association bias (cultural); provided cultural parallel dataset; explored tugging effect between language script bias and cultural instruction.
Neural Dynamic Time Warping Ensemble – Senior Capstone Project (Aug 2023 – May 2024)
Investigated robustness of Neural DTW for time-series classification; implemented ensemble architecture to reduce variance; observed modest stability improvements.

Let's reproduce GPT-2 (124M) – Reproduced GPT-2 from scratch following Karpathy. Added Swiglu, hand-implemented Muon (improved coefficients), RoPE, DeepSeek attention. Turned it into MLM with bidirectional attention → diffusion language model → masked-autoencoder design for text (no advantage found).
Diffusion & flow-matching – Participated in graduate lab journal club; implemented Transition Matching (scalable generative modeling) – a novel ODE-of-velocity approach.

Languages: Thai (native), English (IELTS 7.0)
Programming: Python, SQL (Postgres), HTML/CSS, PHP
Frameworks: PyTorch, JAX, Transformers (Hugging Face), FastAPI, Gradio, Flask
Machine Learning: LLM fine-tuning (QLoRA), MLLMs, OCR, YOLO, metric learning (DLBKL, SP, bias analysis), diffusion & flow-matching, sub-quadratic attention, interpretability
Tools: Git, Docker, AWS, SLURM, VS Code
Libraries: pandas, NumPy, Matplotlib, Plotly, TensorFlow, Keras, scikit-learn