Zhang Yukui / Ktao

Enterprise AI application implementation and requirements alignment (embedded on-site)

• Owned the WeCom AI customer service system: broke consultations into a four-layer collaboration flow of static knowledge, dynamic data, automated actions, and human takeover; it is already in production and continuously iterated.
• Production reliability hardening: hard gates for message deduplication and delivery verification, process supervision and health self-checks, a failure alerting bus, and isolated multi-machine deployment for production and development.
• Quality engineering: maintained 600+ regression tests (real-corpus replay, multi-turn conversation scripts, negative sample battery); set a code-level human review hard gate for outbound mass messaging, so nothing reaches customers without review.
• Built operations data automation: automatic data pulling, business metric definitions governance, real-time dashboards and scheduled report delivery, with gold-standard validation to prevent 'the program ran successfully' from being mistaken for 'the data is correct'; used daily by management.
• Delivered cross-platform lead synchronization and content matrix automation, running stably in production over the long term.

Training mentor for 'AI Implementation in Practice' (contracted)

• Drawing on first-hand production practice with AI customer service and operations data automation, I provide AI implementation training and sharing for merchants; related public materials can be verified on the media page of my personal site.

• All 3 PRs submitted to a 40,000+ star open-source repository were merged (#599 / #600 / #601; total +5464/-150, 57 files); the overall proposal issue was closed as completed by the maintainer.
• Added 9 medical publishing policy goals and reworked AI usage disclosure into a fail-closed process that stops output when information is incomplete; after identifying the mistaken premise that 'Crossref is only one DOI registration agency,' implemented a Chinese literature parsing client and API protocol documentation; added condensed CARE / STARD 2015 / TRIPOD+AI guidelines for the deep research module.
• All three PRs were merged after multiple rounds of adversarial review by the maintainer; the implementation and each round of revisions were completed using AI-assisted coding tools.

• A neuroscience-inspired long-term memory system for AI agents, open-sourced on GitHub with 70 stars / 9 forks — the most recognized of my public works.
• A Python codebase built by directing Claude Code in natural language, implementing memory write, retrieval, and organization mechanisms for agents.

• Turned 3700+ AI collaboration sessions into a structured knowledge base: decision logs, pitfall checklists, and a methodology wiki, so any new agent can inherit full working context within 30 minutes.
• Multi-device session auto-archiving pipeline: cross-machine collection, desensitization, and incremental sync.