PBY

Introducing CU-ICU — my custom-built language model designed specifically for Thai critical care.

Over the past few months, I’ve been pouring my heart into a personal research journey: building CU-ICU, an instruction-finetuned language model purpose-built for doctors working in Thai Intensive Care Units (ICUs).

At this year’s 14th Critical Care Conference (TSCCM 2025), I had the honor of being selected as one of only eight oral presenters—and the only non-MD in the group. The rest were experienced physicians presenting powerful case studies. I walked into the room knowing I was the odd one out... but I was determined to make every word matter.

CU-ICU is based on T5, fine-tuned using cutting-edge lightweight methods like LoRA, AdaLoRA, and IA3. It’s designed to reason through clinical prompts—like diagnosing sepsis or shock—and respond based on guidelines like the Surviving Sepsis Campaign (2021). But more than just medical accuracy, I wanted CU-ICU to be interpretable, fast, and usable in real-life Thai hospital settings.

I designed CU-ICU to be an AI assistant that understands the realities of Thai ICUs, where decisions need to be fast, actionable, and explainable. Doctors shouldn't waste time retyping or searching for textbook information—they should get help that speaks their language and fits their workflow.

If you’re curious about the model, the conference, or why I think this is just the beginning of AI for Thai healthcare, I’ve written it all up here.

📖 Full blog post with slides, paper, and behind-the-scenes, I wrote it all up here (with pics and feels):

CU-ICU: Customizing Instruction-Tuned Language Models for Critical Care | Teerapong Panboonyuen Fine-tuning FLAN-T5 with parameter-efficient methods to adapt LLMs for high-stakes ICU tasks like sepsis detection, mortality prediction, and clinical note generation.

🚀🌍 A New Era of Multimodal Intelligence in Remote Sensing 🌍🚀
After diving deep into the cutting-edge world of Vision-Language Models, I’m excited to share my latest blog post! I’ve just finished writing about how these models are opening up a whole new world of possibilities for remote sensing and geospatial analysis.
By combining visual data with language processing, Vision-Language Models are taking our understanding of satellite imagery, aerial photos, and other forms of remote sensing data to unprecedented heights. The integration of multimodal information has the potential to transform industries ranging from environmental monitoring to disaster response, urban planning, and beyond.
In this post, I break down:
- What Vision-Language Models are and how they work
- Why they're so effective in multimodal tasks
- Real-world applications in remote sensing
- And what does this mean for the future of AI in geospatial tech
It’s an exciting time to be part of this journey, and I can’t wait to see where these advancements take us!
👉 Check out my latest post here: https://kaopanboonyuen.github.io/blog/2025-07-05-visionlanguage-models-for-remote-sensing-a-new-era-of-multimodal-understanding/
Would love to hear your thoughts on how you see these models shaping the future of AI and remote sensing.

Vision–Language Models for Remote Sensing: A New Era of Multimodal Understanding | Teerapong Panboonyuen Exploring how LLMs can enhance remote sensing applications like LULC and image classification using Sentinel, Landsat-8, and THEOS imagery.

🎨✨ AI with Style: ALBERT vs. SLICK — MARSAIL’s New Fashion Statement
https://kaopanboonyuen.github.io/blog/2025-07-02-albert-vs-slick-marsail-new-ai-fashion/
Ever wondered what happens when AI isn’t just smart, but downright stylish? Our newest post dives into the contrasting aesthetic philosophies behind two architectural stars in the MARSAIL universe:
- ALBERT: Sleek, minimalist, and lightning-fast—optimized for edge deployments where efficiency is everything.
- SLICK: Bold, expressive, and rich in spatial reasoning—built to reveal the nuanced patterns in human driving behavior.

Think of it as AI fashion week—two distinct “looks” emerging from our lab’s runway of neural designs. These aren’t just technical choices—they reflect PBY Lab’s core belief: building AI that is both meticulously designed and deeply impactful.
This is where theory meets real-world design. It’s AI that doesn’t just learn—it adapts, reasons, and resonates.
Across industries, we’re seeing AI break into the fashion world—tools that generate haute couture, predict creative trends. In that context, ALBERT and SLICK aren’t just models—they’re statements: that intelligent systems can have personality and purpose.
🔥 Why this matters:
- It’s about designing intelligence that’s both elegant and efficient.
- It’s about using architecture to tell a story—how AI perceives, reasons, and performs.
- It’s about crafting systems with intent, where every layer echoes a design philosophy.
Proud to be part of this creative synthesis with the brilliant team at PBY Lab. If you’re intrigued by how model choices can shape AI’s “style and soul,” this read is for you.

ALBERT vs SLICK: MARSAIL’s New AI Fashion for Real-Time Car Insurance and Garages | Teerapong Panboonyuen Discover how MARSAIL’s teacher–student model duo ALBERT and SLICK revolutionizes car damage detection with unmatched precision and blazing speed.

🚗✨ Introducing MARSAIL: The Smart Engine Behind the Future of Car Insurance
🔗
https://kaopanboonyuen.github.io/blog/2025-07-01-marsail-the-smart-engine-behind-the-future-of-car-insurance/
We’re living in an era where AI is no longer confined to the lab—it’s out on the road, learning in real-time, and starting to understand how we drive. I’m incredibly proud to share our latest work: MARSAIL — a real-time driving intelligence engine that could reshape the entire foundation of car insurance.
MARSAIL goes beyond tracking; it thinks.
By combining deep learning, sensor fusion, edge computing, and self-supervised representation learning, MARSAIL captures not just “how far” you drive, but how well. Smooth turns, safe braking, contextual awareness—MARSAIL learns your patterns and translates them into insights for risk-aware, usage-based car insurance.
This project represents something bigger than just a model or a dataset. It’s about designing AI systems that earn trust, deliver real-world impact, and improve lives daily. Insurance that’s fairer, smarter, and personalized—not based on assumptions, but on actual behavior.
I’m deeply grateful to be part of this journey with the amazing minds at PBY Lab, where vision, theory, and engineering come together with purpose. This is what happens when foundational AI meets domain-specific problems—and starts solving them.
The future of intelligent mobility is already being built. We’re just getting started.

MARSAIL: The Smart Engine Powering the Future of Car Insurance and Intelligent Garages | Teerapong Panboonyuen Explore how MARSAIL, Thailand’s premier automotive AI lab, revolutionizes car insurance and vehicle repair with deep learning.

🏃‍♂️ 42.195K at Chombueng Marathon 2025 🌅
I can hardly put into words how special this marathon was. It was my second full marathon, but this one had an extra layer of meaning because it took place in my hometown of Ratchaburi—the place where it all began for me. The entire experience felt like a beautiful journey of connection and personal growth. 🌳
The race started at 3 AM, a surreal time when the world is still asleep, but the excitement in the air was electric. As the first light of dawn broke over the horizon, I found myself running through the most breathtaking views—rolling mountains, quiet cassava fields, and the peaceful serenity that only a sunrise can bring. Every step felt like a small victory, every stride bringing me closer to my goal. 🌄
What made this marathon truly unforgettable, though, was the support from the local community. Everywhere I ran, I was greeted by cheering voices, encouraging smiles, and the warmth of people who were rooting for me. Their energy was contagious—it made the miles pass by faster and reminded me why running is so much more than just a race. It’s about community, perseverance, and pushing yourself beyond limits you never thought possible. 🙏
I crossed the finish line at 5:36, exhausted but elated. It wasn’t just about the time on the clock—it was about the journey I had just completed, both physically and emotionally. Running through my hometown, revisiting familiar roads and paths, while experiencing it all in such an intense way, gave me a deeper sense of connection to the place that shaped who I am today. 🏅🌟
This marathon taught me so much about resilience, endurance, and gratitude. It reminded me that every step, no matter how challenging, is a part of the bigger picture. I’ll carry this experience with me for a long time.
Check out the full recap of my experience here: https://kaopanboonyuen.github.io/blog/2025-01-20-chom-bueng-marathon-2025/
Thank you to everyone who cheered me on, who believed in me, and who made this race so incredibly memorable. I’ll be looking forward to more races, more challenges, and more moments like these.

🌟 An Incredible Week at GYSS2025 in Singapore! 🌿

What an unforgettable week it’s been at the Global Young Scientists Summit 2025 (GYSS2025) in Singapore! I’m still processing everything that happened—so many incredible moments that truly left me in awe.
I had the honor of presenting my research to Her Royal Highness Princess Maha Chakri Sirindhorn. To have the opportunity to share my work with someone so distinguished, while also receiving valuable feedback, was truly humbling and motivating. The Princess is not only a patron of science but also a champion of innovation, and I’m so grateful for this once-in-a-lifetime chance to connect.
But that’s not all—this week, I was also fortunate to learn from some of the brightest minds in science, including Prof. Adi Shamir (a Turing Award winner) and a group of Nobel Prize and Turing Award laureates. To be in the same room with such brilliant individuals, to hear their insights, and to exchange ideas… It’s an experience that words can barely capture.
The sessions were deeply thought-provoking, sparking conversations that went beyond just academic discussions. The atmosphere was filled with a palpable sense of curiosity and collaboration. It was a gathering of passionate, forward-thinking individuals, all striving to push the boundaries of what we know and what’s possible in science, technology, and society.
I’m truly grateful for the chance to be a part of this incredible gathering, and I’ve walked away from the summit feeling more inspired and motivated than ever. The experiences and connections I’ve made here will guide and shape the work I do moving forward. 🌍✨
This week has been a milestone in my journey—one that I’ll cherish for the rest of my career. Thank you to everyone who made GYSS2025 such a life-changing experience.

Global Young Scientists Summit (GYSS) 2025: Where Science Meets Inspiration | Teerapong Panboonyuen Competing for a prestigious scholarship to attend GYSS 2025 in Singapore. The summit will be an excellent opportunity to connect with global experts and young scientists, fostering innovation and interdisciplinary collaboration.

🏃‍♂️💨 Just Finished an Epic Running Tour Through Osaka, Kyoto, and Tokyo! 🇯🇵✨
What an unforgettable experience! I just wrapped up an incredible running tour through Osaka, Kyoto, and Tokyo, and I’m still buzzing from all the amazing memories. From the moment I laced up my sneakers, it was a whirlwind of racing and sightseeing, all rolled into one!
I hit sub-1-hour times in each city, but what made it even more incredible was the fact that I got to explore these iconic cities while pushing myself to the limit. Each run was like a mini adventure, with the opportunity to experience the best of Japan’s culture, architecture, and city vibes—all at full speed! 🏙️✨
In Osaka, I cruised through neon-lit streets and crossed bustling intersections, feeling the energy of the city while taking in the sights and sounds. From there, it was off to Kyoto, where I ran through peaceful temples and lush gardens, enjoying a slower, more reflective pace as I soaked in the ancient beauty of the city. And then, of course, Tokyo—a vibrant metropolis where tradition meets modernity, and every corner has something new to discover. 🏯🌆
The best part? Racing while exploring! I could never have imagined how exhilarating it would be to run through such diverse environments, whether it was the tranquil streets of Kyoto or the fast-paced urban jungles of Osaka and Tokyo. It felt like two worlds colliding: racing with purpose while immersing myself in the heart of Japan’s rich history and modern flair. 🏃‍♂️💨
I can honestly say this was one of the most *fun* and *rewarding* running experiences I’ve ever had. It wasn’t just about hitting fast times—it was about making memories, discovering new places, and living fully in the moment. Every mile was a chance to connect with the culture, challenge myself, and fall even more in love with this beautiful country. 💖
If you ever get the chance to run through Japan, I highly recommend it. It’s a perfect blend of athletic challenge and cultural immersion.
Can't wait to do it again soon! 🙌

Sub1 Urban Running Experience: Crushing Times and Exploring Osaka, Kyoto, and Tokyo | Teerapong Panboonyuen An incredible running experience across Japan's vibrant cities—Osaka, Kyoto, and Tokyo. I managed to hit sub-1-hour times while soaking in the unique vibes of each place. It was the perfect mix of pushing my pace and enjoying the sights along the way!

🚀 Curious About Fine-Tuning a Satellite-Specific LLM Model? 🌍
If you’ve been exploring the fascinating world of Large Language Models (LLMs) and their potential in geospatial science, this post is for you! I just released a new blog that dives deep into how to fine-tune and deploy an LLM specifically tailored for satellite imagery and geospatial data.
Whether you’re working with Landsat, Sentinel, or other remote sensing data, satellite-specific LLMs can unlock a whole new level of precision and insight. But how do you get started with fine-tuning these powerful models to suit your needs? 🤔 In this post, I break down:
- Key strategies for fine-tuning LLMs on geospatial datasets
- Steps to deploy your model for real-world applications
- The importance of model selection and training for satellite imagery
- Practical insights for geoscientists and AI engineers working in this space
If you’re passionate about AI and its applications in remote sensing, geography, or environmental monitoring, this post offers all the guidance you need to build and deploy a powerful satellite-specific LLM model.

How to Fine-Tune and Deploy a Satellite-Specific LLM Model for Satellite Images | Teerapong Panboonyuen Fine-tuning a Large Language Model (LLM) for satellite imagery requires a deep understanding of both theoretical concepts and practical techniques. This blog post will guide you through the process, incorporating mathematical formulations and theoretical explanations to provide a comprehensive view....

🚗 Exciting News! Our Paper on Road Asset Detection Accepted for Oral Presentation at iCHE 2024! 🎉
I’m thrilled to announce that our paper, “Refined Generalized Focal Loss for Road Asset Detection on Thai Highways”, has been accepted for oral presentation at iCHE 2024! This is a huge milestone for us, and I’m excited to share how we’re using advanced vision models to tackle the challenges of road asset detection on Thai highways. 🌍🚧
In this blog, we dive deep into how Refined Generalized Focal Loss (RGFL) can improve accuracy and robustness in detecting road assets like signs, barriers, and pavement markings—crucial elements for road safety and infrastructure management.
I can’t wait to present our findings and discuss how AI and deep learning are revolutionizing computer vision for real-world applications.

Refined Generalized Focal Loss for Road Asset Detection on Thai Highways Using Vision-Based Detection and Segmentation Models | Teerapong Panboonyuen We are pleased to announce that our paper, titled 'Enhanced YOLOv8-Based Object Detection of Road Assets Utilizing Generalized Focal Loss A Case Study on Thai Highway Imagery', has been accepted for oral presentation at the 5th International Conference on Highway Engineering (iCHE 2024).

🚀 Curious About the Journey to Becoming an AI Research Scientist? 🌟
If you’ve ever wondered what it takes to build a career in AI research, I’ve got something special for you! In my latest blog post, I dive deep into the career paths for aspiring researchers, offering insights and guidance from my recent talk at NSTDA’s One Day Camp held at Sirindhorn Science Home. 🎓✨
Becoming an AI research scientist isn’t just about technical skills—it’s about curiosity, persistence, and passion for solving complex problems. In the blog, I share:
- Key skills and mindsets that help you thrive in research
- My personal journey and lessons learned along the way
- Tips on how to navigate the world of academia and industry, and which opportunities to pursue
- Valuable insights from the brilliant minds I met at the One Day Camp, who shared their own research journeys and career advice.
If you’re passionate about AI, deep learning, or cutting-edge research, this post is packed with actionable tips and motivation to help you carve out your own path in the field.
Don’t miss out on these practical insights—whether you’re just starting or already on your research journey, there’s something in there for everyone!

Career Paths for AI Research Scientists: My Journey and Insights | Teerapong Panboonyuen A thorough examination of my journey as an AI Research Scientist, including insights into balancing academia and industry, significant publications, and guidance for future researchers.

🍄 Thrilled to Unveil MARS: A Groundbreaking Approach to Attention Refinement with Sequential Quadtree Nodes! 🚀
I'm excited to share MARS—our innovative approach that combines Attention Refinement with Sequential Quadtree Nodes to revolutionize how we tackle complex problems in AI and computer vision! This new method takes model efficiency and accuracy to the next level, offering a smarter way to process and analyze visual data. 🌐👁️
MARS is designed to enhance how deep learning models focus on important features in images, making them more effective and scalable. Whether you’re working with high-resolution imagery or complex data, MARS offers a unique, cutting-edge solution.
🔍 Dive deeper with the following resources:
- Research Paper: http://link.springer.com/chapter/10.1007/978-3-031-51023-6_3
- Code: http://github.com/kaopanboonyuen/MARS
- Project Overview: http://kaopanboonyuen.github.io/MARS/
I’m incredibly proud of this work and can't wait for the community to explore it. If you're into AI, deep learning, or computer vision, MARS is a project you don’t want to miss!