AI Product Manager · Developer Tools Builder · Payments & Trust / Safety
Projects · Product context · Principles · 中文
I build AI products that can enter real workflows - not just impressive demos.
- Reliable coding agents - durable sessions, isolated worktrees, recoverable execution, and reviewable changes.
- Agent guardrails - budget controls, loop detection, risky-command protection, and human approval where it matters.
- Verification-first workflows - tests, visual regression checks, and evidence that a claimed outcome really happened.
- Local-first products - keep source code, credentials, sessions, and evidence on the user's machine whenever possible.
My AI product work is grounded in the operational systems around the model:
| Area | Experience I bring |
|---|---|
| Payments & marketplaces | Payment routing, transaction flows, cross-border/local payment integration, and conversion-oriented product design. |
| Trust & Safety | Content-risk policy, model-assisted detection, governance workflows, risk operations, and post-incident feedback loops. |
| AI SaaS & developer tools | 0-to-1 product definition, UX, full-stack delivery, pricing/entitlements, and evidence-aware AI workflows. |
Previously, I worked as an AI Product Manager on payments at Xiaomi and on content Trust & Safety at Tantan. I now use those lessons to build AI tools where control, accountability, and product usefulness are designed together.
| Project | What it does | Why it matters |
|---|---|---|
| RepoReady | Checks whether a repository is ready for Codex, Claude Code, Cursor, and human contributors. | A fast, no-server entry point to safer AI-assisted development. |
| Kimi Code Agent | A native macOS coding-agent workspace with isolated worktrees, auditable tools, MCP / Skills / Hooks, and resumable execution. | Turns agent work into a reviewable local engineering process. |
| DeepSeek Code | A local-first macOS agent harness for conversation, tools, approvals, evidence, and delivery state. | Makes verification status explicit instead of overstating experimental features. |
| agent-guard | Runtime protection for coding agents: loop detection, quota gates, checkpoints, and supervised runs. | Helps prevent runaway loops before they spend the budget or damage the workspace. |
| kimi-boost | One command to add agent skills, reviewer agents, safety hooks, and engineering conventions to a repository. | Makes a strong AI coding workflow repeatable across teams and stacks. |
| VLM-Diff | Hybrid DOM + VLM visual regression testing with inspectable evidence. | Verifies UI changes deterministically first and uses model calls only when needed. |
Model proposes → Policy constrains → Tools execute → Verification checks → Humans approve
This is how I translate payments and safety experience into AI products:
- Separate implemented and verified capabilities from experiments and roadmap ideas.
- Prefer least-privilege access, local storage, isolated worktrees, and reversible actions.
- Treat logs, tests, screenshots, receipts, and diffs as product features - not afterthoughts.
- Design the failure path first: pause, recover, explain, and let the user stay in control.
- Open Support Agent Spec (OSAS) - an open, schema-first contract for governed customer-support agents.
- greenbump - dependency upgrades that hand broken code to an AI repair loop until real builds and tests pass.
- buttonprobe - detecting and repairing dead buttons in local React applications.
TypeScript · Swift / SwiftUI · Tauri · Node.js · React · MCP · CLI tools · GitHub Actions
- X: @shidesheng0218
- 小红书: 海豚号角Delphic (
DolphinHorn_218)
I welcome collaboration on practical AI developer tools, local-first desktop software, and governed agent workflows.
If one of these projects is useful to you, open an issue, start a discussion, or send a pull request.
我是 Mike Shi，一名处在 AI Agent、开发者工具、支付与信任安全 交叉点的产品构建者。
我曾在小米做支付产品，也做过亿级社区的内容安全与用户生态治理；这些经历让我更关注 AI 产品真正进入业务后的问题：权限如何收敛、风险如何控制、结果如何验证、失败后如何恢复，以及用户如何始终保有最终决策权。
我正在构建的产品主线是：让 AI Coding Agent 更可靠、更安全、更容易在真实团队的工程流程中落地。
- 默认本地优先，尽量让代码、密钥、会话与验证证据留在用户设备上；
- 默认可审计，让每一次工具调用、改动、测试和交付都有迹可循；
- 默认受控，让模型提出方案，策略约束行为，验证确认结果，用户保留最终决定权；
- 默认诚实地区分“已实现且验证”与“实验中 / 路线图”。
Build useful things. Show the evidence. Keep people in control.