I am currently pursuing my M.S. degree at Nanjing University, with research and project experience in computer vision, multimodal AIGC, LLM fine-tuning, and AI for Science.

My recent work focuses on building practical AI systems, including e-commerce image generation, community rule violation detection, UAV image regression, medical image classification, and molecular machine learning.

• 🎓 M.S. student at Nanjing University, AI for Science direction
• 🏅 Kaggle Expert with 2 Silver Medals and 1 Bronze Medal
• 🔬 Research experience in equivariant graph neural networks and machine-learning potentials
• 🤖 Project experience in LLM fine-tuning, computer vision, AIGC image generation, and multimodal systems
• 🛠️ Interested in building reliable AI systems that combine algorithms, engineering, and real-world applications

• Kaggle Expert: 2 Silver Medals, 1 Bronze Medal, global ranking Top 1.18%
• Google Jigsaw Rules Classification: Kaggle Silver Medal, ranking 42 / 2445
• CSIRO Image2Biomass Prediction: Kaggle Silver Medal, ranking 83 / 3803
• Jittor Medical Image Classification: Finalist in the Jittor Algorithm Challenge
• E-commerce AIGC Try-on System: built an end-to-end virtual try-on and product image generation pipeline
• FiLM-Δ-PaiNN: first-author research on equivariant graph neural networks for molecular potential prediction

Built an end-to-end virtual try-on and product image generation prototype for e-commerce scenarios.

Tech Stack: FLUX.1 Fill, Gemini Image Generation, SigLIP, LoRA, ComfyUI, FastAPI, LangGraph

Main Contributions:

• Designed a dual-route generation pipeline combining local controllable diffusion models and high-level image generation models
• Used SigLIP-based visual condition injection to preserve garment texture and visual details
• Applied LoRA-based identity customization to improve face consistency in generated product images
• Built an automated workflow with FastAPI, ComfyUI, and LangGraph for generation routing, quality checking, and retry control
• Explored an algorithmic solution to reduce product image production time from traditional shooting workflows to rapid AI-assisted generation

Kaggle Silver Medal, ranking 42 / 2445.

This project focused on community rule violation detection, where the model needed to judge whether a post violated a specific community rule.

Tech Stack: Qwen3-4B, QLoRA, DeepSpeed ZeRO-2, vLLM, GTE, Triplet Loss, DeBERTa-v3, FGM, EMA, Rank Blending

Main Contributions:

• Reformulated the task as rule-conditioned violation detection instead of simple rule ID classification
• Fine-tuned Qwen3-4B with QLoRA under limited GPU resources
• Used DeepSpeed ZeRO-2 and bf16 mixed precision to improve training efficiency
• Built a GTE-based metric learning branch to improve generalization to unseen rules
• Built a DeBERTa-v3 discriminative classifier with FGM adversarial training and EMA
• Combined LLM, metric-learning, and discriminative branches through rank-based blending
• Improved final AUC to 0.929

Kaggle Silver Medal, ranking 83 / 3803.

This project focused on multi-target biomass prediction from UAV top-view grassland images under a small-sample setting.

Tech Stack: PyTorch, timm, DINOv3 ViT-Huge, SigLIP, LightGBM, CatBoost, PCA, GMM, AMP

Main Contributions:

• Designed a DINOv3-based dual-view image feature extraction pipeline for 2000×1000 UAV images
• Split each image into left and right views and encoded them with shared-weight visual backbones
• Developed a Gated Local Token Mixer to improve cross-view token interaction and local texture modeling
• Used structured multi-head outputs to explicitly predict Green, Dead, and Clover biomass, then reconstruct GDM and Total biomass
• Added a SigLIP + LightGBM / CatBoost semantic compensation branch
• Improved weighted R² from baseline 0.54 to 0.63

Finalist in the Jittor Algorithm Challenge, B leaderboard ranking 7th.

This project focused on fine-grained BI-RADS classification from breast ultrasound images.

Tech Stack: Jittor, EfficientNetV2-S, Multi-Dropout, BN-Neck, 5-Fold CV, TTA, AMP, Albumentations, OpenCV

Main Contributions:

• Designed an EfficientNetV2-based fine-grained medical image classification model
• Combined multi-scale features from intermediate and final backbone layers
• Used BN-Neck normalization to stabilize feature representation
• Built a Multi-Dropout classification head with multiple dropout rates and averaged logits
• Used 5-fold cross validation, test-time augmentation, checkpoint ensemble, and mixed-precision training
• Improved robustness under severe class imbalance, especially for high-risk categories with very limited samples

First-author research project on molecular machine learning and AI for Science.

Tech Stack: PaiNN, FiLM, Δ-learning, PyTorch, ASE, DeepMD, equivariant GNNs

Main Contributions:

• Designed an equivariant graph neural network for high-accuracy molecular potential prediction
• Combined PaiNN-style equivariant message passing with FiLM-based physical information modulation
• Used Δ-learning to reduce the learning difficulty between low-fidelity and high-fidelity energy labels
• Evaluated the model on molecular and periodic benchmark datasets
• Achieved significant error reduction compared with direct-learning baselines and several mainstream equivariant models

Research project on quantum-accuracy simulation of hydrogen behavior at mineral interfaces.

Tech Stack: VASP, DeepMD, DPA, LAMMPS, ASE, Python

Main Contributions:

• Built a closed-loop workflow from DFT calculations to machine-learning potential training and molecular dynamics simulation
• Generated and curated more than 40k atomic structures for model training
• Trained machine-learning potentials for hydrogen-water-mineral interface systems
• Used LAMMPS and DeepMD to study interfacial adsorption, diffusion, and structural stability
• Improved simulation efficiency while maintaining near-DFT-level energy and force accuracy

• Python, C/C++
• PyTorch, Jittor, scikit-learn, timm
• FastAPI, LangGraph, ComfyUI
• OpenCV, Albumentations
• Linux, Git, VS Code Remote SSH

• DINOv3
• SigLIP
• EfficientNetV2
• Vision Transformer
• Image classification
• Visual regression
• Multi-scale feature fusion
• Test-time augmentation and model ensemble

• Transformer, BERT, GPT
• Qwen
• QLoRA
• DeepSpeed
• vLLM
• DeBERTa
• GTE embedding models
• Rank blending and ensemble learning

• Diffusion models
• FLUX.1 Fill
• Gemini image generation
• LoRA fine-tuning
• Virtual try-on
• Image editing workflow design
• Generation quality control and automatic retry

• Molecular machine learning
• Equivariant graph neural networks
• PaiNN
• DeepMD
• DPA
• Δ-learning
• Molecular dynamics simulation
• DFT-to-ML potential workflows

I am currently interested in:

• Multimodal AIGC systems for e-commerce and content generation
• LLM fine-tuning and efficient inference
• Computer vision models for small-sample and fine-grained recognition
• Reliable AI workflows with automatic evaluation and retry mechanisms
• Equivariant graph neural networks and machine-learning potentials for molecular simulation

• Email: 1196973334@qq.com
• GitHub: https://github.com/mayin0902