智能体协作Agent collaboration
设计角色分工、审阅门和交接流程，让多个 AI 智能体围绕同一目标稳定协作。Design roles, review gates, and handoffs so multiple AI agents can work toward one goal reliably.
AI 应用 · 智能体协作 · 自动化APPLIED AI · AGENT WORKFLOWS · AUTOMATION
把 AI 接入真实工作流，做成可以运行的自动化项目。 Connecting AI to real workflows—and turning it into working automation.
聚焦 AI 应用、智能体协作与自动化实践，同时具备轻量 Web 开发和部署能力。正在寻找 AI 应用 / 自动化相关实习机会。 Focused on applied AI, agent collaboration, and automation, with hands-on lightweight web delivery and deployment experience. Currently seeking an internship in applied AI or automation.
- AI Agent Workflows
- Python Automation
- ADB / Browser Workflows
- Lightweight Web Delivery
What I bring
不仅会使用工具，更重视交付结果。Beyond using tools, I focus on useful outcomes.
我正在成长为一名能连接需求、AI 工具与实际交付的实践者。面对陌生问题，我会先拆解目标，再选择合适工具，记录过程并持续迭代。 I am growing into a practitioner who connects real needs, AI tools, and practical delivery. When facing an unfamiliar problem, I clarify the goal, choose the right tools, document the process, and iterate.
轻量 Web 落地Lightweight web delivery
把想法整理为可访问、可维护的网页成果，兼顾响应式、可读性和部署流程。Turn ideas into accessible, maintainable web outcomes with responsive layouts and practical deployment.
调研与文档Research & documentation
检索、筛选并组织信息，把复杂过程沉淀为能复查、能复用的文档和工作流。Retrieve, evaluate, and organize information into reviewable, reusable documentation and workflows.
Selected work
三个代表项目，三条真实证据链。Three featured projects. Three traceable proof chains.
chat2api-sft-toolkit
已实践 · 代码公开Practiced · Public code
把吃灰的 Chat 订阅额度，变成 SFT 数据集。Turn idle Chat subscription quota into an SFT dataset.
订阅额度每月用不完，数据标注又贵。项目围绕开源 chat2api，把订阅额度变成可编程的 OpenAI 兼容 API，再用批量问答生成加数据清洗，做出可微调的指令数据集。Subscription quota goes unused every month while data labeling is expensive. Built around the open-source chat2api project, this toolkit turns quota into a programmable OpenAI-compatible API, then ships batch Q&A generation plus data cleaning into a fine-tunable instruction dataset.
- 我的角色My role
- 围绕开源 chat2api 设计批量生成与 SFT 清洗脚本（纯 Python 标准库），基于实测沉淀逆向接线四道坑与 tool_calls 硬提醒。Designed batch-generation and SFT-cleaning scripts (Python standard library only) around open-source chat2api; documented four wiring pitfalls and the tool_calls caveat from real testing.
- 验证结果Verified result
- 断点续跑、指数退避重试均实测；不可达端点返回退出码 1 且不写假数据；去重与过滤统计公开。Resumable runs and exponential-backoff retries verified; unreachable endpoints exit 1 with no fabricated data; dedup and filter stats are public.
- 部署Deploy
- 生成Generate
- 清洗Clean
- 数据集Dataset
- 微调Fine-tune
agent-dispatch-kit
已实践 · 方法公开Practiced · Public method
一个入口，调度你所有的 AI 智能体。One entry point to dispatch all your AI agents.
多个智能体各干各的，没有统一的派发、跟踪与验收。项目给出三层架构与结构化调度/汇报提示词模板，把“派发→跟踪→验收→建议”做成可复用的工作流。Multiple agents working with no unified way to dispatch, track, and verify. This kit provides a three-layer architecture plus structured dispatch/report prompt templates, turning “dispatch → track → accept → advise” into a reusable workflow.
- 我的角色My role
- 设计三层架构与提示词模板，定义验收标准与独立第二模型打分机制，并用真实的 X 内容发布链路验证。Designed the three-layer architecture and prompt templates, defined acceptance criteria with independent second-model scoring, and validated it on a real X publishing pipeline.
- 验证结果Verified result
- 统一调度全流程跑通并发布；独立打分不通过则打回重做；双语文档与验收脚本公开。The full dispatch pipeline ran end-to-end and shipped; independent scoring sends failures back for rework; bilingual docs and the acceptance script are public.
- 派发Dispatch
- 跟踪Track
- 验收Accept
- 打分Score
- 建议Advise
agent-gateway
已实践 · 代码公开Practiced · Public code
把只会聊天的网页号池，变成会调工具的 agent 后端。Turn a chat-only web pool into a tool-calling agent backend.
网页逆向的 ChatGPT 池会静默丢弃 tools 参数，agent 客户端连上也调不了工具。自研协议桥坐在号池上层：一次 XML 伪协议把请求翻成纯文本，深预填充逼模型吐出可解析的工具调用，白名单校验后再回填成标准 tool_use；拒答或乱吐时修复梯度换 nonce 重发，救不回就诚实报错。Web-reverse ChatGPT pools silently drop the tools array, so agent clients can’t call tools. This self-built protocol bridge sits atop the pool: a one-shot XML pseudo-protocol translates the request into plain text, deep prefill coerces the model into emitting a parseable tool call, whitelist validation renders it back as a standard tool_use; on refusal or garbage, a repair ladder retries with nonce rotation and fails honestly when unrecoverable.
- 我的角色My role
- 设计 XML 伪协议与修复梯度，定义 P1–P5 五阶段验收门槛；三轮 A/B 跑批定位三处根因（拒答检测漏判弯引号、上游空正文、strict 清单不对齐），把工具轮合规率从 90% 修到 98.3%。Designed the XML pseudo-protocol and repair ladder, defined the P1–P5 acceptance gates; three A/B batch rounds pinned down three root causes (refusal detector missing curly quotes, empty upstream bodies, strict-list misalignment), lifting tool-round compliance from 90% to 98.3%.
- 验证结果Verified result
- P4 四门槛全过：工具轮 59/60、多轮 20/20、闲聊 0 误触、伪造调用 0；桥审计、客户端日志、执行数据库三向对账零差异。All four P4 gates passed: 59/60 tool rounds, 20/20 multi-turn, 0 chat misfires, 0 forged calls; bridge audit, client logs, and the execution database reconciled with zero drift.
- 伪协议Pseudo-protocol
- 剪枝Prune
- 预填充Prefill
- 校验Validate
- 回填Render
aihot-welfare
已实践 · 补丁公开Practiced · Patch public
把 AI 热点站框架，改成专盯羊毛的福利雷达。Turn an AI hotspot framework into a promo-hunting welfare radar.
开源 AIHOT 原版是新闻监控：采集→判重→双次打分→摘要→聚簇→日报。我要的是福利监控——在它的精选评分里新增 promo_welfare 内容类型，权重向证据强度倾斜：没官方来源、没领取方式的“福利” cred≤3、总分压到 40 以下；过期、名额满的 act≤3；真实可领的按真实价值打分。The open-source AIHOT is a news monitor: collect → dedupe → dual scoring → summarize → cluster → daily digest. I needed a welfare monitor instead — so I added a promo_welfare content type to its selection scoring, weighted toward evidence: promos with no official source or claim path get cred≤3 and a total cap of 40; expired or full-quota ones get act≤3; genuinely claimable promos score at face value.
- 我的角色My role
- 设计 promo_welfare 评分维度与压分规则，给出零成本模型接线（OpenAI 兼容 API key 接入）与低配 VPS Docker 部署记录。Designed the promo_welfare scoring dimension and suppression rules; published a zero-cost model wiring (OpenAI-compatible API key) plus low-spec VPS Docker deployment notes.
- 验证结果Verified result
- 评分补丁与部署配置已公开；AIHOT 负责上游粗筛聚簇，终审（福利优先→爆火潜力→官方核验）由 AI 助手按固定节奏推送。The scoring patch and deployment config are public; AIHOT handles upstream triage and clustering, while final review (welfare-first → viral potential → official verification) is pushed by the AI assistant on a fixed cadence.
- 采集Collect
- 打分Score
- 压分Suppress
- 终审Review
- 推送Push
How I work
先把问题想清楚，再让工具发挥作用。Clarify the problem before putting tools to work.
拆解需求Clarify
先写清问题、公开边界和验收标准；指标无法核实时，改用可复查的定性结论。Define the problem, public boundaries, and acceptance criteria; replace unverifiable metrics with reviewable qualitative claims.
组织协作Orchestrate
用写入声明、角色分工、交接文件和人工审阅门，避免多个智能体互相覆盖。Use write declarations, role boundaries, handoff files, and human review gates to prevent agents from overwriting one another.
交付与复盘Deliver
交付后复查内容、链接、隐私和部署状态，把故障原因与最终取舍沉淀为文档。After delivery, verify content, links, privacy, and deployment state, then document failures and final trade-offs.
工作之外，我持续进行写作、短视频创作、健身和英语学习。这些长期习惯帮助我保持表达力、执行力与持续输入。 Outside project work, I keep practicing writing, short-form content creation, fitness, and English—long-term habits that strengthen communication, execution, and continuous learning.
Learning archive
持续把实践变成可复查的档案。Turning ongoing practice into a reviewable archive.
-
2026-09
chat2api-sft-toolkit 开源：把订阅额度变成 SFT 数据集chat2api-sft-toolkit open-sourced: turning quota into an SFT dataset
批量问答生成、SFT 数据清洗脚本与逆向接线踩坑长文一次公开；双语 README 加懒人一键指令，配套两篇 X 帖子发布。Batch Q&A generation, SFT cleaning scripts, and a pitfalls deep-dive published together; bilingual READMEs plus a one-click lazy prompt, shipped with two X threads.
-
2026-09
agent-dispatch-kit 开源：一个入口调度所有智能体agent-dispatch-kit open-sourced: one entry point for all agents
三层架构、结构化调度/汇报提示词模板与验收脚本公开；派发→跟踪→验收→建议工作流已在真实 X 内容发布链路跑通。Three-layer architecture, structured dispatch/report prompt templates, and the acceptance script published; the dispatch → track → accept → advise workflow validated on a real X publishing pipeline.
-
2026-09
agent-gateway 开源：给没有 function calling 的号池加上工具调用agent-gateway open-sourced: tool calling for pools without function calling
协议桥代码、METHOD 方法手册与 104 项回归测试一次公开；P4 验收三段弧线（90%→98.3%）与三处根因修复实录。Bridge code, the METHOD playbook, and 104 regression tests published together; the full P4 arc (90%→98.3%) with three root-cause fixes documented.
-
2026-09
aihot-welfare 开源：把 AIHOT 改成福利雷达aihot-welfare open-sourced: AIHOT as a welfare radar
promo_welfare 评分补丁、零成本模型接线与低配 VPS 部署记录一次公开；原版做新闻监控，魔改版专盯真羊毛、压制假福利。The promo_welfare scoring patch, zero-cost model wiring, and low-spec VPS deployment notes published together; the base project monitors news, the mod hunts real promos and suppresses fake ones.
-
持续
作品集随实践持续扩充Portfolio keeps growing
仓库保持开源，项目随真实实践逐步增加，每个都力求可运行、可说明、可复用。The repo stays open; projects are added as real practice grows—each one runnable, explainable, and reusable.
最后更新：2026 年 9 月 · 以真实项目持续成长Last updated: September 2026 · learning through real projects
Open to opportunities
正在寻找 AI 应用、智能体自动化方向实习。Seeking an internship in applied AI and agent automation.
也愿意承担轻量 Web 开发与部署工作。欢迎通过邮件交流具体岗位、项目或合作机会。Also open to lightweight web development and deployment work. Email me about a specific role, project, or collaboration.