I turn ambiguous workflows into products that can be used, inspected, and verified—not just feature lists or AI demos.

My flagship work is Keyloom, an AI-powered cross-border B2B lead generation platform for China-based companies expanding overseas. I took it from product positioning to a live product with 3,000+ cumulative users, helping export, sales, growth, and marketing teams move from a target-market hypothesis to verified companies, decision-makers, contact paths, and prioritized opportunities.

My wider portfolio spans a privacy-first Chrome product, an enterprise AI coding workspace, an OpenCode-based maker workbench, a global short-drama MVP, and a safety layer for browser agents. Across them I own the product decision, interaction contract, agent-assisted execution, engineering evidence, and honest delivery boundary.

An AI lead-intelligence and customer-acquisition platform for Chinese exporters and cross-border B2B teams entering overseas markets. Keyloom turns a sales team's target-market definition into verified, ranked, and actionable account and contact leads.

Live product · Service catalog · Documentation · Pricing

• Primary customer: China-based exporters, foreign-trade companies, and cross-border B2B teams whose sales, growth, or marketing staff need to find and qualify overseas buyers without stitching together multiple databases and spreadsheets.
• Core job to be done: define the target market, product, region, company profile, and buyer role; discover matching overseas accounts; identify decision-makers; enrich and verify contact paths; surface commercial signals; and produce a structured lead list ready for human review and outreach.
• Product insight: the bottleneck is not access to more company records. It is turning fragmented, stale, and weakly sourced data into a repeatable answer to three sales questions: Who should we pursue? Why now? How can we reach the right person?
• Product decision: Keyloom is neither a generic AI chat tool nor a raw lead database. AI coordinates multi-source research, entity matching, verification, evidence collection, and ranking, while the salesperson retains control over qualification and outreach.
• Acquisition workflow: Define ICP → Discover overseas accounts → Match and deduplicate entities → Identify decision-makers → Enrich and verify contacts → Rank with business signals and evidence → Review and activate leads.
• Actionable output: each result is designed to carry company and contact context, match rationale, source evidence, confidence or risk signals, and a next-step angle—not just an unqualified email list.
• What I owned: ICP and product positioning; overseas lead-generation journey; roadmap and service prioritization; input and output contracts; account, contact, and signal workflows; Console, Playground, API Key, Usage, Jobs, billing, and documentation; and the operating boundary between AI automation and human sales judgment.
• Supporting platform: a unified capability gateway connects search, public-web intelligence, company and contact data, and specialist APIs behind consistent authentication, synchronous invocation, asynchronous Jobs, structured results, usage metering, and credit-based billing. These are infrastructure for acquisition outcomes, not the headline product.
• Commercial outcome: independently driven from 0→1 to a live product with 3,000+ cumulative users, demonstrating product judgment, developer experience, commercialization, growth, and operational ownership.
• Current boundary: signals indicate relevance and timing; they do not guarantee purchase intent. Keyloom supports evidence-led qualification rather than automated spam or black-box sales decisions. Its production source and provider integrations are private, so the public evidence is the live product and user journey—not an open-source claim.

A privacy-first Chrome focus timer that captures only the tabs a user explicitly selects and turns a session into a local receipt.

• Decision: useful focus context did not require browser-history access, host permissions, page content, or cloud storage.
• Delivery: live Chrome Web Store listing, plus a production-site source and build whose Vercel deployment is currently paused.
• Evidence: 46 automated tests, manifest checks, production build, permission audit, and CI.

An independent multi-user AI coding workspace prototype with one isolated OpenVSCode environment per employee, plus an access portal, model gateway, usage budgets, and secret controls.

• Decision: enterprise adoption needed isolation, cost visibility, and operations controls before another chat UI.
• Delivery: Dockerized workspace lifecycle, per-user state, gateway routing, quotas, secrets, Nginx, and deployment runbooks.
• Evidence: 10 Node source-contract checks and 18 Python tests; upstream OpenVSCode ownership is explicitly separated.

An OpenCode-based AI product workbench that helps non-technical makers begin with product intent, see agent progress, and move toward a presentable artifact.

• Decision: beginners need orientation and confidence before raw agent flexibility.
• Delivery: visual prompt workshop, hosted projects, local-folder sync, progress narrative, operator console, desktop packaging, and production guards.
• Evidence: 127 downstream commits after the recorded import baseline, focused app/desktop tests, executable UX contracts, and exact upstream provenance.

A localized short-drama discovery and viewing product with account, membership, episode, and operations flows.

• Decision: validate the complete viewer journey and content-operations model before pretending a licensed catalog existed.
• Delivery: responsive discovery, title/episode pages, watch flow, account and membership surfaces, i18n, and a Cloudflare worker boundary.
• Evidence: 31 tests, lint, worker type-check, and a 505-page production build. The paused public deployment is disclosed in the repository.

A privacy-aware workflow recorder and deterministic playback core for AI-driven browser automation.

• Decision: an agent-generated workflow must be typed, reviewable, and secret-safe before it can control a browser.
• Delivery: explicit recording sessions, password/token redaction, URL credential cleanup, human handoff, and an injected browser-adapter boundary.
• Evidence: 5 behavior tests, strict TypeScript, clean-install CI, zero high-severity dependency findings, and a zero-finding secret scan.

Ambiguous problem
  → product judgment and non-goals
  → user-facing contract and architecture
  → agent-assisted implementation
  → tests, security checks, and real build
  → honest demo, ownership, and operating limits

I am especially interested in AI products where the difficult work is not “adding a model,” but designing the collaboration between people, agents, tools, data, and business outcomes.

我是李刚，AI 产品经理与 Agent-native 产品实践者。我擅长把模糊需求拆成可验证的产品判断、交互契约和交付证据，并使用 AI Agent 协同完成从 0 到 1 的产品、工程、测试与上线准备。

**Keyloom 是我最重要的代表项目。**它是一款面向中国出海企业、外贸公司和跨境 B2B 团队的 AI 海外线索与获客平台，帮助销售、增长与市场人员完成“定义目标客户—发现海外企业—匹配关键决策人—补全并验证联系方式—识别商业信号—生成可触达线索”的完整链路。

我独立推进 Keyloom 从 0 到 1，负责 ICP 与产品定位、海外获客流程、服务优先级、输入输出契约、企业与联系人数据工作流，以及 Console、Playground、API Key、Usage、Jobs、计费和开发文档。产品通过 AI 编排多源检索、实体匹配、数据清洗、去重、联系方式验证、证据归因和线索排序，把过去依赖多个数据库、搜索引擎和表格的人工研究过程，转化为可复用、可计量、可由人工复核的获客工作流，累计用户 3,000+。

Keyloom 的关键产品判断是：海外 B2B 获客缺少的不是更多名单，而是回答三个问题的可靠证据——**应该联系谁、为什么是现在、怎样找到正确的人。**因此产品不把不透明的“意向分”包装成购买承诺，也不自动进行垃圾群发；AI 负责研究、匹配、验证和排序，销售人员保留最终筛选与触达决策。

这里的六个代表产品分别体现了：商业化数据与工具平台、隐私与产品取舍、企业 AI 平台设计、开源项目二次产品化、全球化内容产品 MVP，以及浏览器 Agent 的安全执行边界。每个案例都明确写出“为什么做、我负责什么、如何运行、怎样验证、当前边界是什么”，方便面试官从任意入口检查我的真实能力。

• GitHub: @lisir202446
• Flagship live product: Keyloom — AI Cross-border B2B Lead Generation
• Shipped Chrome product: Focus Receipt on the Chrome Web Store