yuki.uix yuki-uix

Kunyu Xu (Yuki)

A terminal session. yuki whoami prints an ASCII fish beside: hi, i'm yuki. I build AI agents, eval systems, and developer tools. yuki background prints: I studied architecture before I wrote code. yuki what-keeps-me-curious prints: the jump from it works to it's useful, building it, testing it, finding out.

agent evaluation · grounded generation · local-first safety · developer experience

I build AI agents, evaluation systems, and developer tools that turn LLM demos into measurable delivery decisions. At Thoughtworks, I work across TypeScript, React, Node.js, Python, and Java/Spring Boot—from prototypes and eval harnesses to production workflows.

`01 / systems, not demos`

🧪 pace-triage-agent

Insurance enquiry triage and reply drafting with a frozen golden set, human review queue, and 2×2 evaluation matrix.

The useful result: every tested configuration missed the safety bars, producing a defensible no-ship decision.

🛰️ RepoCoach

A source-learning agent with constructive grounding, complete exit-path safety gates, and measured agent-loop economics.

Validated: agent-loop economics and exit-path safety across 638 tests and 22 merged PRs, with the findings turned into reusable engineering guidance.

🧭 RAG Generative UI Explorer

Grounded evidence becomes schema-constrained, interactive knowledge cards—not arbitrary model-generated UI code.

Measured: dense retrieval reached 63.7% Recall@10 versus 47.6% for BM25; fusion did not beat the best single retriever.

⚡ agent-cost-lab

A measurement lab for the uncomfortable fact that fewer tokens can still mean a larger bill when cache economics change.

First result: one compaction paid back after 18–19 turns, not the 2–4 turns predicted before measurement.

agent-eval-harness • pr-review-agent • beforeshare

`02 / measured impact`

≈90% less time
_{multi-market rollout setup}

1,018 KB → 92 KB
_{evaluation worker bundle}

31 HTTP tests
_{integration coverage}

Top 3 · APAC
_{Thoughtworks AI/works challenge}

Built a project-wide AI code review platform that runs on every PR across multiple TypeScript repositories; it caught a runtime crash missed by ESLint and AI-generated tests.
Built and piloted an agent-configuration evaluation engine with two delivery teams, including a human-in-the-loop improvement cycle.
Encoded multi-market rollout knowledge into a reusable coding-agent workflow, reducing setup from roughly one week to half a day.

`03 / latest writing`

Fresh notes on AI delivery, design engineering, and the parts of software that only become visible after the demo works.

2026-08-27 — We Can Build an Agent in Ten Minutes — Why Does Shipping It Still Take Weeks?
2026-05-31 — Reading baoyu-skills Source Code Through a Design Systems Lens — Three Familiar Patterns
2026-05-28 — I Thought All 9 Token Styles Were Valid — Until I Changed the IA

Read the full archive →

`04 / contribution graph`

`05 / toolbox`

AI engineering: Agent evaluation · RAG / embeddings · MCP · LLM APIs · Prompt / Skill Engineering
Product engineering: TypeScript · React · Next.js · Node.js · Hono · Python · Java · Spring Boot
Delivery: Azure DevOps · CI/CD · Google Cloud · Playwright · Honeycomb

Why architecture still matters to my engineering

I studied architecture before I wrote code. It trained me to see software as a system people move through—not just a set of screens—and still shapes how I design agent interactions, failure paths, and developer tools.

_{A little tree, grown from my GitHub history. Tended daily.}

Portfolio · Juejin · LinkedIn · Dev.to

Navigation Menu