Hi, I'm Kunyu Xu.
AI Engineering · Agent Systems · Developer ToolsChengdu, China · Remote
I build ideas into things people can use.
Software Engineer at Thoughtworks, working across TypeScript, React, Node.js, Python, and Java/Spring Boot. I build agents, evaluation harnesses, and production workflows with clear evidence for delivery decisions.
Systems, evidence, and delivery
Building e-commerce SaaS at Thoughtworks while developing agent evaluation, AI code review, and reusable AI delivery practices across teams.
Tech Stack
I came to software through architecture and UX. That background trained me to read products as flows: where people hesitate, where state gets lost, and where a system quietly stops helping.
Today I work across frontend, BFF, and backend services. I treat agents as part of that delivery stack, with interfaces, APIs, data contracts, observability, and explicit failure paths.
My projects use frozen test sets, safety gates, human review, and cost measurements to make model behaviour inspectable.
I want to know what works, what fails, and whether the evidence supports the next delivery decision.
Focus Areas
Work History
Software Developer
Work spans React and TypeScript purchase flows, BFFs, Java and Spring Boot services, and Kafka event pipelines. I have also owned features from technical design and analytics through production deployment.
Built a project-wide AI code review platform that runs on every pull request across multiple TypeScript repositories. It caught a runtime crash missed by ESLint and AI-generated tests.
Built an agent-configuration evaluation engine and human-feedback improvement loop, then piloted it with two delivery teams.
Encoded multi-market rollout knowledge into a reusable coding-agent workflow, reducing setup from roughly one week to half a day.
Measured outcomes
Selected AI Systems & Experiments
◇ EXPERIMENT
AI COST · MEASUREMENT
A measurement lab for real agent bills. The first experiment found that one context compaction took 18–19 turns to pay back, far beyond the 2–4 turns predicted before measurement; fewer tokens did not automatically mean lower cost.
● LIVE
AI AGENT · E-COMMERCE
Where filters fall short, agents begin. An AI shopping agent built with Claude API and tool calling — demonstrating how natural-language intent understanding replaces multi-level filter UX.
● LIVE
CLAUDE CODE · SKILL · VISUALIZATION
You wrote a great article — most people won't click the link. A Claude Code `/viz` skill that detects article structure and generates mind maps, interactive HTML, and shareable image cards — each output links back to the source so every audience finds their entry point.
◇ EXPERIMENT
AI AGENT · EVALUATION · SAFETY
Insurance enquiry triage and reply drafting with a frozen golden set, human review queue, and 2×2 model/prompt evaluation matrix. Every tested configuration missed the safety bars, producing a defensible no-ship decision.
◇ EXPERIMENT
AI AGENT · COST & EVALUATION
A source-learning agent used to test agent-loop economics, bidirectional safety gates, and cross-process state. The findings from 638 tests and 22 merged pull requests became reusable engineering guidance.
◇ EXPERIMENT
RAG · GENERATIVE UI · EVALUATION
Grounded evidence becomes schema-constrained, interactive knowledge cards instead of arbitrary model-generated UI code. Across 60 evaluation questions, dense retrieval reached 63.7% Recall@10 versus 47.6% for BM25.
◇ EXPERIMENT
AI COST · MEASUREMENT
A measurement lab for real agent bills. The first experiment found that one context compaction took 18–19 turns to pay back, far beyond the 2–4 turns predicted before measurement; fewer tokens did not automatically mean lower cost.
● LIVE
AI AGENT · E-COMMERCE
Where filters fall short, agents begin. An AI shopping agent built with Claude API and tool calling — demonstrating how natural-language intent understanding replaces multi-level filter UX.
● LIVE
CLAUDE CODE · SKILL · VISUALIZATION
You wrote a great article — most people won't click the link. A Claude Code `/viz` skill that detects article structure and generates mind maps, interactive HTML, and shareable image cards — each output links back to the source so every audience finds their entry point.
◇ EXPERIMENT
AI AGENT · EVALUATION · SAFETY
Insurance enquiry triage and reply drafting with a frozen golden set, human review queue, and 2×2 model/prompt evaluation matrix. Every tested configuration missed the safety bars, producing a defensible no-ship decision.
◇ EXPERIMENT
AI AGENT · COST & EVALUATION
A source-learning agent used to test agent-loop economics, bidirectional safety gates, and cross-process state. The findings from 638 tests and 22 merged pull requests became reusable engineering guidance.
◇ EXPERIMENT
RAG · GENERATIVE UI · EVALUATION
Grounded evidence becomes schema-constrained, interactive knowledge cards instead of arbitrary model-generated UI code. Across 60 evaluation questions, dense retrieval reached 63.7% Recall@10 versus 47.6% for BM25.
Featured Articles & Notes
Everyone says building an agent has never been faster; everyone also says shipping one is brutally hard. Both are true — they describe different things. This piece unpacks the weeks between demo and delivery: discovering how it fails, defining what "correct" means, and re-verifying after every change. The cost didn't drop, it moved from writing to verifying — and got more expensive on the way.
Nine token profiles, one HTML structure — each presenting a completely different product personality. They all seemed valid, until I asked: what if this weren't a landing page? Move Playful into a data dashboard, drop AI Gradient into a payment flow — the wrongness is immediate.
The functionality works, but the interface is lacking. We get AI to optimize it, but after a dozen rounds of revisions, it just gets messier. This article documents my shift in understanding from "the prompt wasn't written well enough" to "there's no design system at all," and how design tokens transform this from something mystical into an actionable definition.
Scroll restoration after a redirect isn't a scroll problem — it's a timing problem. State must survive the jump, the element must appear in the DOM, and it must be fully painted — in that order, at the right moment in the browser rendering pipeline.
Not low-ranking — completely invisible to Google. A micro-frontend SEO post-mortem and a four-layer diagnostic framework: link discovery, accessibility, content rendering, semantic understanding. Fix the wrong layer and nothing changes.