仓库/Skill 库

完全在浏览器中通过 WebAssembly 实现文档转换、检查、OCR 与翻译。保留版式的翻译，全程本地、离线优先。Convert, inspect, OCR, and translate any document entirely in the browser via WebAssembly. Layout-preserving translation, fully local, offline-first.

面向长生命周期 Agent 与有状态负载的会话中心计算：暂停即保留记忆，按需恢复。Session-centric compute for long-lived agents and stateful workloads. Pause with memory intact; resume on demand.

面向有据可循的 Agent 工作流的 TypeScript 参考实现，支持混合检索、人工审核与可回放的审计追踪。TypeScript reference implementation for grounded agent workflows with hybrid retrieval, human review, and replayable audit trails.

面向 GitHub Copilot 的 Agentic 开发生命周期——计划、工作、证据与教训全部沉淀在 GitHub 上，而非聊天窗口中。Agentic development lifecycle for GitHub Copilot — the plan, the work, the evidence, and the lessons all live on GitHub, not in chat windows.

KARZOUN-X —— 在地球通信延迟或不可用条件下，面向航天器自主故障诊断的资源感知、安全门控本地 AI 开源研究。KARZOUN-X — Open research on resource-aware, safety-gated local AI for autonomous spacecraft fault diagnosis under delayed or unavailable Earth communication.

Astro 静态站点，英意双语，安全设计（security-by-design），AI 辅助的月度 RAG 杂志。部署于 Cloudflare。Astro (static), bilingual EN/IT, security-by-design, AI-assisted monthly RAG magazine. Deployed on Cloudflare.

驾驶 Claude Code Agent 的驾驶舱——打磨「指示→等待→审查→再指示」循环的桌面 Cockpit（非官方）。A cockpit for piloting Claude Code agents — 指示→待機→レビュー→再指示のループを磨くデスクトップコックピット (unofficial)

探索高达 400B 参数的先进稀疏 Mixture-of-Experts 模型，采用创新训练技术实现卓越性能🚀 Explore advanced sparse Mixture-of-Experts models with up to 400B parameters, featuring innovative training techniques for superior performance.

Microwave Method：agent 工厂 + 受治理的 LLM wiki。每个 agent 一次性重型处理，每个 feature 轻量处理。上下文只"烹饪"一次，按缓存价"回热"。Microwave Method: agent factory + governed LLM wiki. Heavy pass once per agent, light pass per feature. Context is cooked once, reheated at cache price.

Claude Code 动态工作流的指挥中心 — 实时 agent 可观测性、跨项目运行历史、保存与重放。零依赖的 Claude Code 插件。Mission control for Claude Code dynamic workflows — live agent observability, cross-project run history, save & re-run. A zero-dependency Claude Code plugin.

软件开发software development

涵盖 RAG、AI Agent、评估与商业智能的实战 AI 应用项目集。Practical AI application projects across RAG, AI agents, evaluation, and business intelligence.

借助这款自托管、本地优先的应用，跨多个账户与多种货币追踪收入和支出。Track income and expenses across multiple accounts and currencies with this self-hosted, local-first application.

🐺 创新型 LLM 驱动 12 人狼人杀模拟，针对直播优化并支持 OBS 无缝集成🐺 Experience an innovative, LLM-driven Werewolf simulation for 12 players, optimized for live streaming and designed for seamless integration with OBS.

基于 n8n 构建的 AI agent，用于阅读、分类和回复询价邮件，并与 Google Sheets 和 Gmail 集成。AI agent created in n8n for reading, sorting, and replying to quote emails, integrated with Google Sheets and Gmail.

Pinakes —— 可移植、Agent 优先的知识库。一个目录即一个 KB。Pinakes — a portable, agent-first knowledge base. One directory = one KB.

用前瞻性奖励训练面向动机式访谈的小型 LLM 治疗师：Preference Tree Optimization (PTO) vs GRPO，配合模拟患者和 LLM 评委。里赫曼大学硕士论文。Training small LLM therapists for Motivational Interviewing with look-ahead rewards: Preference Tree Optimization (PTO) vs GRPO, with simulated patients and LLM judges. Master's thesis, Reichman University.

大语言模型监督微调的反馈对齐方法Feedback alignment for supervised fine-tuning of large language models

metaScreener——基于插件的桌面应用，用于 human-in-the-loop systematic literature screening。在顺序可审计的 pipeline 中结合确定性启发式过滤与 LLM 推理，通过 SHA-256 校验包实现完全可复现。MIT 协议。metaScreener — a plugin-based desktop application for human-in-the-loop systematic literature screening. Combines deterministic heuristic filters with LLM inference in a sequential, auditable pipeline. SHA-256 verified bundles for full reproducibility. MIT licensed.

多阶段气候声明检索与排序：BM25、稠密 ANN、融合、重排序与评估Multi-stage climate claim retrieval and ranking: BM25, dense ANN, fusion, reranking, and evaluation

由机器强制执行的 AI Agent 前端 UI/UX Skill 包。注册表 + 懒加载：2,018 token 的路由在每次请求中从 334k token 深度内容中按需加载一个 Skill。包含 11 个发布阻塞闸门，其中一项用于校验 Skill 包自身文档与参考资料是否遵循其规则。Machine-enforced frontend UI/UX skill pack for AI agents. Registry + lazy loading: a 2,018-token router loads one skill per request out of 334k tokens of depth. 11 release-blocking gates, including one that checks the pack's own docs and reference material against its own rules.

Telegram RAG 机器人，以暗黑荒诞的《权力的游戏》谋士风格回答问题；支持英中双语，基于 LLaMA 3.3 70B + Firestore 向量搜索。A Telegram RAG bot that answers your questions as a darkly absurdist Game of Thrones advisor — bilingual (EN/ZH), powered by LLaMA 3.3 70B + Firestore vector search.

这个仓库展示了我的技能、项目，以及从数据工程转型为数据科学家角色的持续学习历程。This repository showcases my skills, projects, and continuous learning journey as I transition from Data Engineering to a Data Scientist role.

🛠 通过此硬件插件提升 vLLM 在 Kunlun XPU 上的性能，无缝集成主流 AI 模型并优化执行效率🛠 Enhance vLLM performance on Kunlun XPU with this hardware plugin, offering seamless integration for popular AI models and optimized execution.

🧪 实验性本地优先 RAG，附带桌面 GUI：将文档拖入文件夹即可对话聊天，内置 Ollama / fastembed / OCR / reranker。🧪 Experimental local-first RAG with desktop GUI — drop documents in a folder, chat with them. Ollama / fastembed / OCR / reranker built in.

部署并运行本地 RAG 实例Deploy and run a local RAG instance

🔍 Self-Corrective RAG 基于 LangGraph 与 Google Gemini 优化查询并评估文档相关性，提升检索效果🔍 Enhance your searches with Self-Corrective RAG, a system that optimizes queries and evaluates document relevance using LangGraph and Google Gemini.

基于 .NET 的 RAG 文档摄取流水线 — 追踪 git 备份仓库中的文件夹，提取并分块文件，保持向量索引同步。.NET document ingestion pipeline for RAG — tracks a folder of files in a git-backed vault, extracts and chunks them, and keeps a vector index in sync.

为 AI agent 提供持久化、结构化记忆。基于 Go 的本地优先 MCP server：语义召回、去重、provenance、冲突解决。单一静态二进制，零依赖。Persistent, structured memory for AI agents. Local-first MCP server in Go: semantic recall, dedupe, provenance, conflict resolution. One static binary, zero dependencies.

多租户 SaaS 自动化平台，具备事件驱动工作流、AI 驱动的文档处理与实时集成。Multi-tenant SaaS automation platform with event-driven workflows, AI-powered document processing, and real-time integrations.

从零开始在 PyTorch 中构建 decoder-only Transformer，涵盖分词、预训练、监督微调，以及基于核心张量操作的对齐。Build decoder-only Transformers from scratch in PyTorch, covering tokenization, pretraining, supervised fine-tuning, and alignment using core tensor operations.

ChatGPT 风格 AI 助手，面向 UET Taxila 学生——通过 RAG 基于真实大学数据回答招生/学费/课程问题，而非猜测。采用 Next.js、Convex 与 multi-LLM 路由。ChatGPT-style AI assistant for UET Taxila students - answers admissions/fees/courses questions using real university data via RAG, not guessing. Next.js, Convex, multi-LLM routing.

使用 Qwen3-TTS 在本地 GPU 上克隆声音并从文本生成语音，提供端到端训练流水线。Clone voices and generate speech from text locally on your GPU using Qwen3-TTS with an end-to-end training pipeline.

保留图表的转换器 —— DOCX/XLSX 转 Markdown，采用原生 OOXML 图表数据提取（无需光栅化/OCR/VLM），并提供零损耗的复合图表标记。Converters where figures survive — DOCX/XLSX to Markdown with native OOXML chart-data extraction (no rasterize/OCR/VLM) and zero-loss composite-figure markers

3D 记忆岛屿，让 Google ADK agent 将你的记忆保存为有据可查的书：基于 Gemini on Vertex AI、Firestore、Pub/Sub 的 dreaming 机制构建知识图谱与 dark keepers，配合 Cloud TTS/STT 语音。All Things Agentic Hackathon 参赛作品。3D memory island where Google ADK agents keep your memories as grounded books: Gemini on Vertex AI, Firestore, Pub/Sub dreaming that builds a knowledge graph and dark keepers, Cloud TTS/STT voice. All Things Agentic Hackathon entry.

基于证据的构建流水线，编排 10 个 AI 子 Agent，并通过确定性 5 道关卡验证系统证明正确性。Evidence-based build pipeline that orchestrates 10 AI sub-agents and proves correctness through a deterministic 5-gate verification system.