HKU BEng AI & Data Science × BA AI and Humanities 港大 AI 与数据科学 × AI 与人文双学位
Building at the intersection of technology × humanities 我相信技术与人文的交汇处，才能诞生真正有意义的作品
Thinking about thinking — question the frame before solving inside it. What assumptions am I making? Is this even the right problem? 思考「思考」本身——先质疑框架，再在框架内解题。我在做什么假设？这真的是对的问题吗？
Map concrete problems to abstract structures. An agent pipeline is function composition; fact-checking is an adversarial verification graph. Find the universal pattern, then specialize. 把具体问题映射为抽象结构。Agent 管线是函数组合，事实核查是对抗验证图。先找通用模式，再特化落地。
Hypothesis-driven, evidence-first. Built a 5-phase fact-check pipeline (SIFT / IFCN / CoVe / FIRE / FABLE); benchmark LLMs systematically. Claims without sources are noise. 假设驱动、证据优先。自建五阶段事实核查管线 (SIFT / IFCN / CoVe / FIRE / FABLE)，系统性评测大模型。没有来源的断言只是噪音。
"The core isn't building — it's researching your way to a good solution, then applying it." 「核心不在于构建，是调研到一个好的方案，然后把这个方案应用上去。」
Decompose, compose, verify. Small verifiable units chained into emergent capability — subagent architectures, pipelines, verification loops. 分解、组合、验证。小而可验证的单元链接成涌现能力——子代理架构、流水线、验证循环。
"If it can be made reusable, make it reusable — write it as a Skill, a Pipeline, a benchmark." 「能做成可复用的东西，就尽量做成可复用的东西——写成 Skills、写成 Pipeline、写成 benchmark。」
Ship fast by working with AI, not against it. Karpathy's 4 rules running in production; subagent → quality gate → verified merge. No completion claim without fresh verification evidence. 与 AI 协作而非对抗。Karpathy 四规则真实在产；子代理 → 质量门 → 验证后合并。没有新鲜的验证证据，就不宣称完成。
"A sub-agent testing loop running all night costs a few dozen RMB in tokens — and saves me hours." 「sub-agent driven testing loop 循环一晚上顶多三四十块钱，但省了我几个小时。」
Unified Fact / Logic / Theory & Tool-Use verification. Two-round architecture: Round 1 can be any generation method; Round 2 is always the same verification pipeline. Bounded Rationality — enforcement via hooks, not model internalization. 统一的事实 / 逻辑 / 理论与工具使用核查。双轮架构：第一轮可以是任何生成方式，第二轮永远走同一条验证管线。有限理性范式——靠 hooks 外部强制，而非指望模型自觉。
"Round 1 optimizes fluency, Round 2 optimizes accuracy — separating them removes the trade-off." 「第一轮优化流畅度，第二轮优化准确率——把两者分开，就消除了取舍。」
General engineering methodology: research → double-check research → plan → fact-check plan → auto-execute (SDD) → deliver → notify owner. Language-agnostic, domain-agnostic. Skip non-owner decisions — surface only what the owner must decide. 通用工程方法论：调研 → 复核调研 → 计划 → 核查计划 → 自动执行 (SDD) → 交付 → 通知 owner。语言无关、领域无关。跳过非 owner 决策，只上报 owner 必须拍板的事。
Unified platform base for ALL DSE products: E2B managed sandbox (Firecracker microVM per-user isolation, ~150ms cold start), OpenCode serve as unified agent backend, Next.js conversational frontend, shared credits wallet (SQLite, product-scoped), launchd + Cloudflare tunnel + Telegram push. ONE agent entry + ONE auth + ONE wallet across products. 所有 DSE 产品的统一平台基座：E2B 托管沙箱（Firecracker microVM 按用户隔离，~150ms 冷启动）、OpenCode serve 统一 Agent 后端、Next.js 对话式前端、共享积分钱包（SQLite，按产品记账）、launchd + Cloudflare tunnel + Telegram 推送。全部产品共享同一 Agent 入口、同一认证、同一钱包。
Tiered research at 3 rigor levels: Quick (vibe research) / Deep (discovery-first, 2+ independent sources, 2+ search backends, hygiene checklist) / Academic (literature → paper → citations → slides). 三档严谨度的分级调研：Quick（快速探路）/ Deep（先探索再分解，2+ 独立信源、2+ 搜索后端、事后卫生清单）/ Academic（文献 → 论文 → 引用 → 幻灯片）。
Tester subagent team tests any product (Web / CLI / desktop / API) from a real user's perspective → structured issues → Fixer subagent team repairs → retest. Loop until no P0/P1. Project-agnostic. Tester 子代理团队从真实用户视角测试任何产品（Web / CLI / 桌面 / API）→ 产出结构化 issue → Fixer 子代理团队修复 → 复测。循环直到没有 P0/P1。项目无关。
Underneath: a skill registry with semver + dependency lock (skill-lint), Keychain-backed secrets, hooks-based enforcement — infrastructure that travels across agents (proven by the AgentDropOne migration). 底座：semver + 依赖锁的技能注册表 (skill-lint)、Keychain 密钥管理、基于 hooks 的强制执行——能跟着 Agent 一起迁移的基础设施（AgentDropOne 已验证）。
OPPO voice assistant replacement — a real agent living on the phone. DeepSeek API + tool calling + system-level integration: the assistant button summons an agent, not a chatbot. 替换 OPPO 语音助手——真正住在手机里的 Agent。DeepSeek API + 工具调用 + 系统级集成：按下助手键唤起的是 Agent，不是聊天机器人。
Systematic benchmark of domestic Chinese LLMs on tool calling and daily tasks — DeepSeek / GLM / Qwen head-to-head. Independent research, no vendor marketing. 国内大模型系统性评测：工具调用与日常任务，DeepSeek / GLM / 通义千问横向对比。独立研究，不依赖厂商宣传。
Personal agent ecosystem: 30+ custom skills, ChromaDB local vector search, MCP servers, cross-platform chat analysis — the era that seeded today's skill library. 个人 Agent 生态：30+ 自研技能、ChromaDB 本地向量检索、MCP 服务器、跨平台聊天分析——今天技能库的起点。
3-mode English speaking practice platform built on ElevenLabs ConvAI Agent, targeting HK schools — with a free tier for SEN and low-income students. 基于 ElevenLabs ConvAI Agent 的三模式英语口语练习平台，面向香港学校——为 SEN 与低收入学生提供免费档。
"Ship the full cycle, fast — walk the whole product journey, validate the business model, even the launch." 「快速交付全产——走完整个产品流程，验证商业模式甚至宣发。」
"Echoes of Thought" — AI interactive visual novel with GenAI as a dynamic Game Master. Analyzes YouTube watch history to surface cognitive biases and echo chambers through personalized stories. 《Echoes of Thought》——以生成式 AI 为动态 Game Master 的互动视觉小说。分析 YouTube 观看历史，用个性化故事让用户看见自己的认知偏差与信息茧房。
Rebuilding DSE education with AI — full-stack platform, automated content production, live product with real users. 用 AI 重塑 DSE 教育——全栈平台、自动化内容生产线，真实上线、真实用户。
BEng in AI & Data Science + BA in AI and Humanities (Philosophy as advanced elective) — engineering rigor and humanistic depth in one curriculum. AI 与数据科学工学士 + AI 与人文文学士（哲学高阶选修）——工程的严谨与人文的深度，在同一张课表里。
5 days, 0 to 1, any track. Looking for a team of 2–4 who ship real products and care about taste. 5 天从 0 到 1，赛道不限。寻找 2–4 人团队：交付过真实产品，在意产品品味。