Hi, I'm
Sr. Staff Engineer & Bar Raiser @ Coupang
When Execution becomes cheap, Judgment is the moat — but Judgment without Ground Truth is just guessing
Learn MoreI am Johnson Li, a Sr. Staff Engineer & Bar Raiser at Coupang, based in Seoul. My work spans systems engineering, mobile infrastructure, and AI engineering, with a focus on architecture, performance, and developer tools. I build Booster and Graphite.
Recently, I have been focused on verification in agentic coding: preserving the original goal across long tasks, turning production scenarios into acceptance criteria, and proving that an optimization works. I also write about AI infrastructure, investing, and independent thinking.
"The ceiling of How depends on the precision of What."
— Johnson
From performance work on Graphite to the control mechanisms behind long-running agents, I keep asking: how do we prove the result meets the original requirements?
Capture real inputs, expected results, and resource limits as test cases, so acceptance criteria survive changes to the implementation.
Study how goals persist across turns, how execution resumes, and what evidence is needed to accept a task as complete.
Examine where harnesses add value, including how business constraints, permissions, and verifiable outcomes enter the system.
Extract business rules and engineering knowledge from code so people and AI can query, understand, and reuse it.
Distinguish continued execution from useful progress, and examine whether each round produces enough feedback to judge the result.
Extend engineering questions into industry research: capacity replication, delivery speed, and constraints on scaling.
Mobile app quality optimization toolkit — performance detection, threading optimization, package slimming, and system bug hot-fixes
Queryable program graphs from JVM bytecode for AI code analysis, with ongoing work on large-scale queries and validation
Opens Claude Code's black box — visualize context composition, chain-of-thought reasoning, tool calls, and compaction summaries
Quantitative US stock analysis — three-factor alpha scoring, capital flow tracking, and market turn detection
Real-time AI vs AI debate arena — orchestrates multi-service battles with streaming bidirectional sync, no API keys needed
Cross-platform app framework powered by SolidJS reactivity, Yoga flexbox, and 2D rendering — no DOM or WebView
Built company-wide mobile infra from scratch, launched R.LUX luxury app from scratch, empowered engineering with AI. Cross-team architecture decisions, root cause analysis of performance bottlenecks, tech debt prioritization; as Bar Raiser, making independent hiring judgments beyond the hiring manager.
"Defining technical direction, upholding talent standards — ensuring every new hire raises the bar."
Large-scale mobile architecture design and performance optimization.
"With hundreds of millions of DAU in short-video, there are no shortcuts to peak experience optimization."
Led driver/rider mobile platform architecture and open-sourced the Booster build optimization tool (5000+ GitHub Stars).
"Solid infrastructure isn't designed — it emerges under real business pressure."
Smart TV SDK & Cloud IDE development.
"Cloud Native wasn't a buzzword here — it was built line by line."
Where it truly began — a solid engineering foundation was forged here.
"IDE, compilers, operating systems — the foundation of systems engineering was laid here."
How production experience becomes acceptance criteria, through work on Graphite
Persistent goals, continuation across turns, and the limits of completion checks
Where engineering value remains as common capabilities move into SDKs
Tracing capacity, delivery, and replication speed through cost curves
The feedback needed to turn continued execution into useful progress
你好，我是
Sr. Staff Engineer & Bar Raiser @ Coupang
当 Execution 变得廉价，Judgment 才是护城河——但没有 Ground Truth 的 Judgment，只是猜测
了解更多我是李景森，Coupang 的 Sr. Staff Engineer & Bar Raiser，现居首尔。从系统工程、移动端基础设施到 AI 工程，我持续在复杂系统中做架构、性能优化与开发工具建设，也是 Booster 和 Graphite 的开发者。
最近，我把更多精力放在 Agentic Coding 的验证问题上：怎样保留长任务的原始目标，怎样把生产场景变成验收标准，怎样证明一次优化真的有效。博客也记录我对 AI 基础设施、投资与独立思考的探索。
"How 的上限，取决于 What 的精度。"
— Johnson
从 Graphite 的性能优化到长任务 Agent 的控制机制，我持续追问：怎样证明结果满足了最初的要求？
把真实输入、预期结果和资源边界固化为 Test Case，让实现可以迭代，验收标准不随代码漂移。
研究跨轮次目标保存、续跑和完成验收，追踪 Agent 如何把原始目标推进到有证据的结果。
持续审视 Harness 的价值：工具编排之外，业务约束、权限和可验证的结果如何进入系统。
提取代码中的业务规则与工程知识，让人与 AI 能够查询、理解和复用。
区分持续运行与有效进展，关注每轮执行之后的反馈是否足以判断方向和结果。
从软件工程延伸到产业研究，关注产能复制、交付速度与规模化约束。
移动应用质量优化套件——性能检测、多线程优化、包体积瘦身、系统 Bug 热修复
将 JVM 字节码转化为可查询的程序图，为 AI Agent 提供代码结构与依赖分析，持续完善大规模查询和验证体系
让 Claude Code 的黑箱变透明——可视化上下文组成、思维链推理、工具调用和压缩摘要
美股量化分析系统——三因子 Alpha 模型选股 + 资金流向追踪 + 市场拐点检测
AI vs AI 实时辩论竞技场——多 AI 服务自动对战，流式双向同步，无需 API Key
跨平台应用框架，基于 SolidJS 响应性 + Yoga 弹性布局 + 2D 渲染，无需 DOM/WebView
从零构建全公司移动端基础设施，从零打造 R.LUX 奢侈品 App，用 AI 赋能工程效率。跨团队架构决策、性能瓶颈根因分析、技术债优先级判断；作为 Bar Raiser 独立于 hiring manager 把关每一个招聘决策。
"定义技术方向，守住人才标准——确保每一个进来的人都能拉高团队的平均水平。"
大规模移动端架构设计与性能优化。
"亿级 DAU 的短视频场景，极致体验优化没有捷径。"
主导司/乘移动端中台架构，开源 Booster 构建优化工具。
"好的基础设施不是被设计出来的，是在真实业务的压力下长出来的。"
Smart TV SDK 与 Cloud IDE 开发。
"Cloud Native 不是概念，是在这里一行一行写出来的。"
职业生涯的真正起点，扎实的工程基础在这里练就。
"IDE、编译器、操作系统——系统工程的底子，是在这里打下的。"