Kieran Zhang

Timeline

1. 2010s
2. 2017

   EducationEnrolled in a finance program.+

   Finance was the family expectation — a 'stable' career. By the second year I was writing Excel macros to backtest paper-trading ideas, getting my first real taste of code's leverage.

3. 2019

   LifeFirst non-trivial Python — a thesis saved by code.+

   My senior thesis needed text analysis on thousands of corporate annual reports. Three sleepless nights with BeautifulSoup, then a week to ship. After that I knew: code is not a tool, it is a stance for solving problems.

4. 2020s
5. 2020

   EducationSelf-taught C/C++, from primer to STL source.+

   Turned down a research-analyst offer and locked myself in a dorm with C++ Primer and the SGI STL source. Three months of barely getting things to compile, then a fourth month where template metaprogramming finally clicked. Two habits stayed: read disassembly on Godbolt, read source before docs.

6. 2021

   RoleJoined Huawei: Cangjie compiler internals.+

   First job, deep end. Worked on the front-end of Cangjie — Huawei's in-house language — covering the parser, type checker, and error recovery. Read more LLVM source in those months than I had written code in my life. The discipline of 'understand the system before changing it' came from here.

7. 2022

   MakingBrought LSP into VS Code and LeetCode editors.+

   Wrote the LSP integration that powered Cangjie's tooling inside VS Code, and a stripped-down version that ran in LeetCode's browser editor. A lot of TypeScript reading. Performance budgets in a browser context taught me to think about latency before correctness.

8. 2023

   MakingAI diagnostic assistant for HarmonyOS app teams.+

   Before HarmonyOS Next launched, I built a diagnostic assistant — knowledge graph plus LLM — that triaged adaptation issues raised by app teams like Feishu and Douyin. Watching senior engineers use it changed how I think about AI: it does not replace experts, it amplifies them.

9. 2024

   LifeLeft Huawei to work independently.+

   No heroic reason — I had seen what I came to see, and the rest of my time felt like it should belong to me. Saved enough to live frugally for a couple of years, and walked out.

10. 2024

    MakingHacked on AutoGLM until the agent could really move.+

    Forked Zhipu's AutoGLM and wired it into macOS system APIs so it could drive a real browser, Cursor, Linear. Two months of work that mostly taught me where current agents actually break: not at the model, at the tool layer.

11. 2025

    MakingGUI agent test runner on Alibaba's MAIUI.+

    Traditional E2E testing relies on hand-written selectors. Inside an agent-driven UI, that contract is broken. I built a layer on MAIUI that lets an agent observe the screen and propose its own test cases. Still iterating, but it now runs against my own products.

12. 2025

    MakingStarted Bubble, an agent that keeps its own diary.+

    Bubble is not just a task agent — it writes its own field notes, maintains a worldview, and carries continuity across sessions. I think of it as a long-term experiment in 'AI as companion, not utility.' Its diary lives on this site.

    • Bubble's diary
13. 2026

    MakingThis site as an ongoing public archive.+

    Engineering notes, AI experiments, photographs, and reading notes — all kept in one place. Not chasing reach, just trying to leave something a future me will recognise as honest work.