๐Ÿ‘‹ Jiaxing Ni (ๅ€ชๅฎถๅ…ด)

Computer Science student at Beijing Institute of Technology, Xu Teli Elite Class.

I build near a small constellation: AI products, evals, agentic RL, multimodal systems, and the engineering needed to make fragile ideas usable.

Good systems feel quiet on the surface. Underneath, there are datasets, prompts, failure cases, retrieval traces, metrics, and a few scripts watching the night shift.

๐Ÿงญ Coordinates

  • Identity: student, builder, product-minded engineer.
  • Base: Beijing / Shanghai.
  • Current focus: LLM/VLM applications, evaluation infrastructure, agent behavior, retrieval-heavy products.
  • Taste: prototype quickly, measure honestly, keep only what survives contact with real edges.

๐Ÿ› ๏ธ What I Build

  • AI products from 0 to 1: user problems, product loops, workflow design, usable demos.
  • Evals and benchmarks: OOD cases, task metrics, A/B experiments, annotation rules, agent-arena style tests.
  • LLM/VLM systems: fine-tuning, OCR/document parsing, multimodal fusion, RAG, context engineering.
  • Agentic RL thinking: reward signals, behavior shaping, online/offline evaluation.
  • Engineering foundations: reproducible experiments, Docker, distributed training/inference, small tools that remove confusion.

๐Ÿ“Œ Proof of Work

AI Travel Planner โ€” Product & Tech Lead

Built a travel planning product around itinerary generation, recommendation, long-context reasoning, and real user loops.

  • 100K users in three months, DAU share over 35%.
  • Recommendation accuracy 82%, user satisfaction 91%, order conversion up 40%.
  • Built a travel-agent benchmark with OOD scenes and long-context consistency checks.
  • Worked with roughly 40B tokens of training data and 5B domain tokens.

AI Note Assistant โ€” Baidu Hackathon

Multimodal note and knowledge retrieval assistant: document parsing, semantic search, personalized summaries, Notion-style organization.

  • Led model selection, retrieval-generation pipeline design, prompt/context engineering, and end-to-end evaluation.
  • Used PaddleOCR-VL, embeddings, retrieval, and Notion integration.
  • Won the Baidu Hackathon Model Power Award and entered incubation.

ByteDance Xpert โ€” Expert Data & Evaluation

Worked on expert-level annotation standards and benchmark design.

  • Designed data quality control flows, labeling guidelines, and evaluation mechanisms.
  • Provided samples and analysis for model alignment and iteration.

PKU Wangxuan Institute โ€” Research Intern

Worked with Prof. Peng Yuxin’s group on fine-grained sports video understanding.

  • 3D pose reconstruction, action localization, model training, data preprocessing, and metric design.

Other Fragments

  • Robot perception and control for quadruped systems.
  • Target detection modules for laser weeding robots.
  • Multimodal human-computer interaction for commercial aerospace scenarios.

๐Ÿงฐ Toolkit

  • Languages: Python, C++.
  • ML: PyTorch, PaddlePaddle, Transformers, DeepSpeed.
  • Retrieval: FAISS, Milvus, reranking, vector indexing.
  • Infra: Docker, Git, distributed training/inference.
  • Open source: Paddle, PaddleOCR, PaddleDetection, and nearby ecosystems.

Tools are not ornaments here. A tool earns its place when the next experiment becomes less vague.

๐ŸŒ Find Me Online

๐Ÿ“ About This Blog

This blog is my field notebook: build logs, experiments, product reflections, and records of what the metric did not explain.

Built with Hugo and PaperMod, hosted on GitHub Pages, and open source on GitHub.