Personal Homepage · CUMT Alethic Insight Lab
Local-first AI for language learning.
Byte-level interfaces for language models.
中国矿业大学本科一年级，CUMT Alethic Insight 实验室核心成员。 一边造能在自己机器上跑起来的外语学习系统，一边追问：分词器真的是必须的么？
Position / 关于
应用侧，我做本地优先的外语学习 AI：词汇、作文批改、实时语音对话、课堂实时助教，敏感数据默认留在本机。 研究侧，我维护 FLUED——一个关于无分词器字节到潜变量接口的公开研究档案，版本化实验、消融和证据注册表全部留档。
On the application side, I build local-first AI for foreign-language learning — vocabulary, essay grading, real-time voice dialogue, and in-class teaching assistants, with sensitive data staying on your own machine by default. On the research side, I maintain FLUED, a public research archive on tokenizer-free byte-to-latent interfaces, with versioned experiments, ablations, and an honest evidence registry.
Directions / 方向
全栈本地优先的外语学习平台：智能词汇、六维作文批改、ASR→LLM→TTS 实时语音、课堂实时助教，端云协同推理。
A full-stack, local-first platform for foreign-language learning: smart vocabulary, six-dimension essay grading, real-time ASR→LLM→TTS voice dialogue, and an in-class assistant, with end-cloud collaborative inference.
AIforLanguageLearning →FLUED：原始字节流能否在进入主干模型之前，被翻译为紧凑、可解码、可逆的潜变量表示？主张什么、不主张什么，全部写在证据注册表里。
FLUED: can a raw byte stream be translated into a compact, decoder-compatible, reversible latent representation before the backbone? What we claim — and what we don't — is all written into the evidence registry.
FLUED →Open WebUI 工具调用生态：把科研调研沉淀为图、树、时间线的 ResearchMap；考试期复习桌面系统 ExamPilot。
An Open WebUI tool-calling ecosystem: ResearchMap, which distils research surveys into graphs, trees, and timelines; and ExamPilot, a minimalist exam-season desktop companion.
ResearchMap →早期工程作品：覆盖建筑、交通、机房、生活的 AI 节能系统（CV / RL / GraphRAG），以及建筑模型到 Three.js 的精确转换工具链。
Earlier engineering work: an AI energy-saving system spanning buildings, traffic, labs, and daily life (CV / RL / GraphRAG), plus an exact building-model to Three.js conversion toolchain.
SmartEnergySavingSystem →Open Work / 项目
Operating System / 工作方式
敏感数据默认留在本机，云端只是可选项而非前提。
Sensitive data stays on your machine by default; the cloud is an option, never a prerequisite.
实验记录、数据来源与判词全部留档，证据先于结论。
Experiment logs, data sources, and verdicts are all archived — evidence before conclusions.
公开 AI 辅助的范围与验证方式，不接受盲目提交，包括我自己。
AI assistance is disclosed with its scope and verification — no blind submissions, including my own.
文档、接口与维护边界预先写明，让别人接得住。
Docs, interfaces, and maintenance boundaries are written upfront, so others can pick the work up.
Contact / 联系