PROJECT INDEX · 2026.08

Projects

Not a list stacked by date — the strongest projects go first. What counts here is whether something is actually running, what problem it solves, and whether the results can be verified externally.

Ranking calibers

Real users and sustained operation take priority, then business results, risk control, and external verification. The ranking changes as project status changes; a lower spot doesn't mean failure — it just means the public evidence or the scope of impact is weaker right now.

Projects ranked by combined evidence

  1. In production

    RANK 01 · Enterprise delivery

    AI customer service system for a medical aesthetics chain on WeCom

    Splits real storefront inquiries into four layers — AI answers, structured queries, automated actions, and human takeover — with process supervision and a hard review gate to control production risk.

    Strongest evidence In continuous use at real storefronts; 600+ regression tests cover real transcripts, multi-turn dialogue, and negative samples.

    AI + human layersDesktop automationProduction hardening
  2. In production

    RANK 02 · Enterprise delivery

    Business data automation: dashboard + report bot

    Aligns metric definitions across intranet reports and manual Excel files, pulls the data automatically, and delivers it to management's phones and business group chats.

    Strongest evidence The data pipeline runs automatically every day; data leaving the system is constrained by permissions and a hard review gate.

    Data pipelineMetric definition governanceWeCom bot
  3. In stable long-term operation

    RANK 03 · Enterprise delivery

    Douyin Laike × Feishu (Lark) lead sync

    Wires up lead pulling, phone-number decryption, ad-data correlation, and Feishu writes, and adds read-back validation and failure alerts.

    Strongest evidence Lead turnaround dropped from days to minutes, and failures no longer stay silent.

    API integrationData syncMonitoring & alerts
  4. Running in a real business

    RANK 04 · Enterprise delivery

    Xiaohongshu (RedNote) multi-account content matrix

    Orchestrates collection, style-based rewriting, image generation, publishing, and comment replies into a multi-account workflow, with humans still owning topic selection and taste.

    Strongest evidence Daily operation of 5 accounts compressed into one Feishu table check per day.

    Content automationPlaywright / CDPFeishu hub
  5. 3 PRs merged

    RANK 05 · Upstream contribution

    Upstream contribution to the Academic Research Skills medical module

    Added medical publishing policies with a fail-closed disclosure process, a Chinese-language literature parsing client, and EQUATOR reporting guideline coverage to an upstream repo with 40,000+ stars; all three PRs were merged after multiple rounds of adversarial review by maintainers.

    Strongest evidence GitHub merge records and the review process are publicly verifiable; the overall proposal issue was closed by maintainers as completed. Upstream stars aren't counted as my own.

    Medicine × AIOpen-source collaborationFail-closed
  6. Personal production system

    RANK 06 · Engineering system

    Bilingual site translation pipeline: quality gates on LLM output

    Maintains the site's English mirror with incremental translation, tag-skeleton comparison, wording red lines, number-loss protection, and a release gate.

    Strongest evidence Pages on this site, blog posts, and structured data are continuously generated through the same quality pipeline.

    LLM engineeringQuality gateZero npm dependencies
  7. 70 star / 9 fork

    RANK 07 · Open-source exploration

    Agent Memory Supersystem

    Inspired by neuroscience, a public implementation of an agent's memory writing, retrieval, and organization mechanisms.

    Strongest evidence The repo, code, and stars/forks are publicly verifiable; earlier performance claims that lack reproducible experimental support are not carried over.

    PythonAgent memoryOpen-source exploration

Only projects that can be clearly explained publicly are shown. Client and internal-system information is anonymized; what I was personally responsible for is still defined by the notes on each project's detail page.