Dai

Dai

he/him

Trying to teach computers to see things

About me

Hey! I’m Dai, a final-year Computer Science student at Hanoi University of Industry. Like many CS students these days, I got caught up in the AI wave that’s basically taking over the world right now.

This blog is where I document my learning journey and the projects I’m working on. If you’re also a student trying to figure out AI and deep learning, or just curious about what I’m up to, feel free to look around!

What I’m into

Where am I

I’m really interested in the intersection of computer vision and deep learning. There’s something cool about teaching computers to “see” and understand images. I also dabble in NLP when I have time.

Most of my coding is in Python (obviously, since it’s basically the AI language), and I know C from my earlier courses. Still figuring out a lot of things as I go!

Why I write here

Honestly, writing about what I’m learning helps me understand it better. Plus, if someone else finds it useful, that’s a bonus! I’m hoping to:

  • Share what I learn: Maybe save someone else the headache of debugging the same issues I ran into
  • Keep track of my progress: It’s nice to look back and see how far I’ve come

What’s next

Right now I’m focused on landing a good internship where I can get some real-world AI experience. After that, who knows? Maybe research, maybe industry - I’m keeping my options open and seeing where this journey takes me.

Current projects

May 2024 – Present

  • Developed a solution for accessing Kaggle via SSH using zrok or ngrok
  • Gained 76 stars and 19 forks on GitHub

BraTSAM (Python, PyTorch, PEFT, Transformers)

Sep 2025 – Present

  • Fine-tuned the Segment Anything Model (SAM) for brain tumor segmentation on MRI scans using LoRA
  • Used parameter-efficient fine-tuning (peft) to create a prompt-based segmentation model
  • Developed a fully automatic pipeline by training a YOLOv8m detector to generate prompts for SAM
  • Surpassed SOTA brain tumor segmentation models (nn-UNet) on BraTS GLI 2024, achieving an average Dice score of 0.8985

Diffmoji (Python, PyTorch, W&B, ComfyUI, CFG)

Sep 2025 – Oct 2025

  • Built a Diffusion model for emoji generation from scratch
  • Implemented a text-to-image Denoising Diffusion Probabilistic Model (DDPM)
  • Designed a U-Net with ResNet blocks, self-attention, and feature modulation conditioned on time and text embeddings
  • Integrated Weights & Biases (W&B) for real-time logging of loss, samples, and system metrics

paddy_doctor (Python, PyTorch, timm, W&B, Gradio)

Dec 2024 – Feb 2025

  • Fine-tuned ResNet26d (from timm) to classify paddy diseases using competition dataset
  • Built a Gradio web app allowing users to upload paddy images for real-time disease prediction
  • Ranked in the top 27% of the competition with 98.6% accuracy on the private leaderboard