Local-first personal data systems, and how much judgment to hand to a language model.
Computer Science and Technology · Hunan Agricultural University · Expected June 2028
Email · GitHub · Zhihu · Bilibili
- Direction — local-first tools for personal records: data stays on the device, and model output is checked against the source.
- Currently — building applied AI agents, personal tools, and scripts; learning cybersecurity engineering; and self-studying the full lifecycle of production-grade LLM agent systems, centered on guardrail ablation experiments.
- Reach me — nekobear.x@outlook.com.
ArkTS · ArkUI · AGPL-3.0
A HarmonyOS application for recording, reviewing, and calibrating judgments. A prediction freezes its confidence, reason, and settlement date on save, so the later review tests the original call instead of a rewritten one. Records sit in an encrypted twelve-table database, and the review agent reads them only after the user enables it.
34 pages · 33 components · 27,399 lines · 453 commits · 7 device test cases
Python · Jinja2 · MIT
An analysis pipeline that reads a diary or essay collection as a dialogue between the author and their self, other people, time, and surroundings. Scripts do the counting, an LLM does the annotation, and a validation gate blocks any report that lacks evidence or verified quote positions before HTML is written.
15 disciplinary lenses · 3,447 lines of Python · 711-line template · 6 test modules
Spring Boot 3.2 · Spring AI · Vue 3.5 · MySQL · Apache-2.0
A software engineering practicum built as a working system: a 2.5D greenhouse heatmap driven by live temperature, a scheduled Spring AI report pushed to online users over WebSocket, and supply-chain records with logical deletes and traceable purchase history.
121 Java files · 29 Vue components · ~12,500 lines · 14 controllers · 16 mappers · 28,000 time-series records
| Project | What it does | Built with |
|---|---|---|
| logisim-ita-lab | Text-first workflow for a circuit lab platform: parse a .circ template, generate a candidate, verify it along one of four routes. Port coordinates were measured from Logisim-ITA 2.16.2.0 rather than assumed. |
Python, Java |
| deepseek-ucs-adapter | Rewrites DeepSeek export fragments into the format a reference parser reads, with deterministic UUIDs; 24 unit tests run in CI. | Python, GitHub Actions |
| userscript-download-lab | Adapter-driven download engine with a fixture site; 83 assertions across nine test files, plus six written experiments. | JavaScript, Node.js |
| transportconsult-master | Dijkstra routing over an editable train and flight network, with fare and travel-time objectives. | C++ |
| Area | What I use |
|---|---|
| Languages | TypeScript, ArkTS, Python, Java, C++, JavaScript, SQL |
| Mobile | HarmonyOS ArkTS, ArkUI state management V2, NavPathStack, relationalStore |
| Backend | Spring Boot 3, MyBatis-Plus, Sa-Token, MySQL, WebSocket, REST |
| Frontend | Vue 3, Vite, Pinia, ECharts, Element Plus |
| Model integration | Spring AI, OpenAI-compatible endpoints, prompt templates, LLM annotation pipelines |
| Engineering | Python and Node test suites, GitHub Actions, Docker, Git, ADRs |
- Zhihu — writing as 曦昼.
- Bilibili — publishing as 曦昼Sunshine, and moderating a community there.


