10+ 年开发经验。崇尚简洁、实用。
Software engineer with 10+ years of experience. I like simple tools, practical systems, and code that does not try too hard to look clever.
这些年做过产品、平台、开发者工具、基础设施、Web3、小爬虫、机器人，也写过很多给自己提效的小工具。最近主要在折腾 LLM 和 AI Agent：记忆、工具调用、浏览器自动化、AI 网关、评测，以及怎么让 Agent 真正变得好用。
I have built products, platforms, developer tools, infra, Web3 bits, crawlers, bots, and plenty of small workflow tools. These days I spend most of my tinkering time on LLMs and AI Agents: memory, tool use, browser automation, gateways, evals, and making agents actually useful.
有些东西想明白了，或者还没完全想明白，我会顺手记下来。主要是分享，也方便未来的自己少踩几次坑。
I write things down when I figure something out, or when I am still figuring it out. Mostly for sharing, and partly so future me has fewer mysteries to solve.
- 让 AI Agent 真的记得住东西，而且记忆可追溯、可重建
Making AI Agent memory traceable, rebuildable, and less hand-wavy. - 搭一些 Agent 基础设施：网关、工具、连接器、运行时、编排
Building agent plumbing: gateways, tools, connectors, runtimes, orchestration. - 做 LLM 代码能力评测，给感觉配一点数据
Running LLM coding evals so the vibes have numbers attached. - 折腾浏览器自动化，尤其是那些必须依赖真实登录态的工作流
Playing with browser automation, especially workflows that need a real logged-in session. - 把突然冒出来的想法快速做成 demo，再看看值不值得继续打磨
Turning random ideas into working demos, then deciding which ones deserve more polish.
| Project | Notes |
|---|---|
| 🧠 memo | 给 AI Agent 做记忆的一次尝试：原始对话可追溯，长期记忆可重建，Agent 也可以自己查证。 My attempt to give AI Agents a real, inspectable memory. |
| 🚦 aigate | AI API 网关，用来管理多个模型 Provider：重试、故障切换、多租户、Token、guardrails、文档。 An AI API gateway for taming multiple providers in one place. |
| 🔀 multicall | 给 coding agents 用的轻量编排内核。有时候一个 Agent 不够，有时候多个 Agent 又需要一个裁判。 A small orchestration kernel for coding agents. |
| 🏭 app-factory | 按 spec 生成 App seed 的小工厂，目标是快速把想法跑起来，但不要每个 App 都长得像模板。 A spec-driven app factory for turning rough ideas into runnable seeds. |
| 📊 llm-benchmark | FastBench，LLM 代码能力评测工具，多 Provider、规则裁判、LLM 裁判、历史记录和报告。 A practical LLM coding benchmark setup with history and reports. |
| 🛡️ safecafe | Safenet staking 的非托管工具，有 Web App、CLI、交易规划和 Safe Transaction Builder 导出。 A non-custodial Safenet staking interface and CLI. |
| ✉️ agentmail | 给 AI Agent 用的邮件 API。让 Agent 有邮箱，但不用真的打开浏览器装成人类。 Email API for AI Agents on Cloudflare Workers. |
| 🌉 browser-rpc-bridge | 驻留在浏览器里的 RPC 桥，可以让脚本借用真实 Chrome 登录态发请求。 A browser-resident RPC bridge for authenticated-session workflows. |
| 🎧 easyVoice | 文本/小说转语音、有声书、字幕、流式播放、多角色配音。 Text and novel-to-speech, audiobooks, subtitles, streaming playback, and multi-character voices. |
| 📈 longbridge-va | 基于长桥 OpenAPI 的 VA 定投 CLI，把策略计算、模拟、回测这些枯燥数学交给工具。 A Value Averaging investment CLI for Longbridge OpenAPI. |
| 🔎 nodex | AI 增强的 Node.js 爬虫平台：队列、代理、提取、存储、索引、监控面板。 An AI-enhanced Node.js crawler platform. |
| 💱 dexflow | DEX 聚合器实验，围绕多链交换、路由比较、钱包集成和实时价格做的一套全栈尝试。 A DEX aggregator experiment around multi-chain swaps and route comparison. |
| 🧪 learn-llm | 我的 LLM 实验场：mini agents、浏览器自动化、GPT/VITS、连接器、沙箱、RAG、embedding。 My LLM playground: mini agents, browser automation, connectors, sandboxes, RAG, embeddings, and notes. |
有些项目暂时未开源。Some projects are not open-source yet.
- GitHub: @cosin2077
- 我会分享一些 LLM、AI Agent、基础设施、产品工程相关的笔记、实验和半成品想法。
I share notes, experiments, and half-finished thoughts around LLMs, AI Agents, infra, and product engineering.