Selected work   /   Research note   /   Collaboration   /   All repositories

I'm a Data Science undergraduate with a research-first approach to building software. My work moves between medical NLP, reliable systems, and small, thoughtful products.

I enjoy taking an idea all the way from an experiment to something others can inspect, reproduce, and use.

Research & evaluation   /   Backend & agent systems   /   Product engineering

01   /   TCM-SD42   _{RESEARCH / MEDICAL NLP}

A controlled comparison of four language models for syndrome differentiation, with reproducibility materials, statistical analysis, and documented limitations. The experiment uses a fixed 42-class task derived from TCM-SD.

_{Python   /   PyTorch   /   Transformers   /   Evaluation}

02   /   PhytoSkill-Spark   _{SYSTEMS / AGENT RUNTIME}

Composable skills for medicinal-plant analysis, with an SDK, registry, execution runtime, verification gates, and repeatable test and evaluation workflows.

_{Python   /   Runtime design   /   Testing   /   Tooling}

03   /   EduMind   _{APPLICATION / EDUCATION}

A personalized learning system connecting learner profiles, retrieval-supported tutoring, learning-path planning, and traceable multi-agent workflows.

_{TypeScript   /   RAG   /   Application engineering}

04   /   Zhiguang-Planting   _{PRODUCT / 3D WEB}

A dental education prototype bringing together spatial visualization, question-answering, and a Spring Boot service layer.

_{Vue 3   /   Three.js   /   Spring Boot}

05   /   daily-zero-decision   _{PRODUCT / INTERACTION}

A gentle habit and action app, with thoughtful interactions, browser-local progress, and a growing visual sky.

_{React   /   TypeScript   /   Product design}

06   /   Qwen2609-Evaluation   _{TOOLING / REPRODUCIBILITY}

Read-only benchmark result aggregation. Preserves the grader's verdicts and refuses to publish a formal rate for partial runs.

_{Python   /   Evaluation tooling   /   Result provenance}

In the TCM-SD42 controlled comparison, the domain-specific encoder ZY-BERT achieved 75.77% accuracy and 66.51% macro-F1 on the clean held-out split (N = 4,218). These results describe the study's historical, fixed TCM-SD-derived 42-class task, not the official normalized 148-class TCM-SD benchmark.

Methods, statistical tests, and reproducibility materials →

I also contribute to shared codebases. A few merged PRs in the collaborator-owned public repository tangjue3/v4:

Frontend and backend APIs · #1   /   Homepage redesign · #7   /   Local deployment, MySQL, RAG · #11

Research: Python / PyTorch / Transformers / LoRA & QLoRA
Systems: FastAPI / retrieval / agent workflows / testing / Docker
Products: TypeScript / React / Vue / Spring Boot / Three.js

_{Build carefully. Measure honestly. Keep learning.}
Explore the work →