About

👋 Hi, I’m Jiaxing Ni (倪家兴)

Welcome to Njx’Log — my personal blog where I document my learning journey in AI and technology.

I’m a LLM & Multimodal Engineer passionate about turning research into reproducible training/inference pipelines. I have hands-on experience in LLM fine-tuning, RAG systems, multimodal fusion, and 0→1 product development.

🔧 What I Do

I specialize in bridging the gap between AI research and production systems:

• 🧠 LLM Fine-tuning: LoRA, QLoRA, SFT, full-finetune with domain-specific data
• 🔍 RAG Systems: FAISS/Milvus vector indexing, retrieval-generation pipelines, rerank strategies
• 🖼️ Multimodal AI: Vision-language models, OCR (PaddleOCR-VL), cross-modal representation learning
• 📊 Evaluation & Annotation: Automated evaluation pipelines, A/B testing, IAA metrics, adversarial test sets
• 🎯 RLHF: PPO, reward modeling, online/offline evaluation

🛠️ Tech Stack

• Frameworks: PyTorch, PaddlePaddle, DeepSpeed, Transformers
• Vector DBs: FAISS, Milvus
• MLOps: Kubernetes, Docker, Git
• Languages: Python ，C++

📚 Education

🎓 Beijing Institute of Technology (北京理工大学)

• Computer Science, Xu Teli Elite Class (徐特立英才班)

💼 Experience & Projects

🤖 AI Note Assistant — Baidu Hackathon 2025.5 - Present

🏆 “Model Power Award” (模力加成奖) & Incubation

Multimodal note-taking & knowledge retrieval assistant with document parsing → vector indexing → personalized summaries.

• Tech: PaddleOCR-VL, sentence-transformers, FAISS, Notion API integration
• Performance: End-to-end response <500ms (parallelization + caching)

✈️ AI Travel Planner — Product & Tech Lead 2024.9 - 2025.6

Personalized travel itinerary generation combining LLM with retrieval and recommendation.

• Tech: RAG pipeline (FAISS), domain fine-tuning (40B tokens), DeepSpeed distributed training (8×A100), rerank + constrained decoding
• Results: 82% recommendation accuracy, 40% conversion rate improvement, 100K users in 3 months

📋 ByteDance Xpert — Annotation & Evaluation System

Led expert-level annotation standards, quality control workflows, and evaluation systems.

• Results: IAA consistency improved >15%, evaluation rules reused in subsequent model iterations

🔬 Research Intern — PKU Wangxuan Institute

3D pose reconstruction and fine-grained action analysis for sports videos under Prof. Peng Yuxin’s lab.

• Participated in model training, data preprocessing, and evaluation metric design

🏆 Awards & Honors

• 🥇 National AI Competition (Intelligent Agent) — First Prize
• 🥇 Challenge Cup National Entrepreneurship Competition — First Prize
• 🥇 National Mathematical Modeling Contest — Beijing First Prize
• 🥈 Douyin Creator Program — National 2nd Place (Innovation/Project)
• 🏅 China Software Cup (Robot Dog Track) — Excellence Award
• 🎓 Baidu Wenxin Mentor — External mentor for enterprise LLM applications

🌐 Find Me Online

• 🐙 GitHub

📝 About This Blog

This blog is:

• Built with Hugo & PaperMod
• Hosted on GitHub Pages
• Open source on GitHub

“The faintest ink is better than the best memory.” — I write to learn, and I share to remember.

Thanks for stopping by! Feel free to explore the posts or reach out if you have any questions.